From a7166feb1053814b7dd27f3879ae38acfc9637fc Mon Sep 17 00:00:00 2001
From: Devin Matthews <dmatthews@utexas.edu>
Date: Mon, 25 Jun 2018 12:09:18 -0500
Subject: [PATCH] Finish macroization of assembly ukernels.

---
 frame/include/bli_x86_asm_macros.h            |   6 +-
 .../3/bli_gemm_bulldozer_asm_d4x6_fma4.c      | 190 +++++++++--------
 kernels/knl/1m/bli_dpackm_knl_asm_24x8.c      |   9 +-
 kernels/knl/1m/bli_spackm_knl_asm_24x16.c     |  14 +-
 kernels/knl/3/bli_dgemm_knl_asm_24x8.c        |   2 +-
 kernels/knl/3/bli_sgemm_knl_asm_24x16.c       |   8 +-
 kernels/penryn/3/bli_gemm_penryn_asm_d4x4.c   |  97 ++++-----
 .../penryn/3/bli_gemmtrsm_l_penryn_asm_d4x4.c |  54 ++---
 .../penryn/3/bli_gemmtrsm_u_penryn_asm_d4x4.c |  51 +++--
 kernels/penryn/3/bli_trsm_l_penryn_asm_d4x4.c |  26 +--
 kernels/penryn/3/bli_trsm_u_penryn_asm_d4x4.c |  26 +--
 .../3/bli_gemm_piledriver_asm_d8x3.c          | 200 +++++++++---------
 .../3/bli_gemm_sandybridge_asm_d8x4.c         | 198 ++++++++---------
 kernels/skx/3/bli_dgemm_skx_asm_16x12_l2.c    |   2 +-
 kernels/skx/3/bli_sgemm_skx_asm_32x12_l2.c    |   2 +-
 kernels/zen/3/bli_gemm_zen_asm_d6x8.c         | 184 ++++++++--------
 kernels/zen/3/bli_gemm_zen_asm_d8x6.c         | 184 ++++++++--------
 kernels/zen/3/bli_gemmtrsm_l_zen_asm_d6x8.c   |  96 ++++-----
 kernels/zen/3/bli_gemmtrsm_u_zen_asm_d6x8.c   |  98 +++++----
 19 files changed, 726 insertions(+), 721 deletions(-)

diff --git a/frame/include/bli_x86_asm_macros.h b/frame/include/bli_x86_asm_macros.h
index 5379b87bd..792dc46c1 100644
--- a/frame/include/bli_x86_asm_macros.h
+++ b/frame/include/bli_x86_asm_macros.h
@@ -98,7 +98,7 @@
 #define COMMENT_BEGIN "#"
 #define COMMENT_END
 
-#define BEGIN_ASM __asm__ volatile (
+#define BEGIN_ASM() __asm__ volatile (
 #define END_ASM(...) __VA_ARGS__ );
 
 
@@ -149,8 +149,8 @@
 
 #endif
 
-#define begin_asm BEGIN_ASM
-#define end_asm END_ASM
+#define begin_asm() BEGIN_ASM()
+#define end_asm(...) END_ASM(__VA_ARGS__)
 
 #define label(...) LABEL(__VA_ARGS__)
 #define imm(...) IMM(__VA_ARGS__)
diff --git a/kernels/bulldozer/3/bli_gemm_bulldozer_asm_d4x6_fma4.c b/kernels/bulldozer/3/bli_gemm_bulldozer_asm_d4x6_fma4.c
index 059d54e64..8b8771226 100644
--- a/kernels/bulldozer/3/bli_gemm_bulldozer_asm_d4x6_fma4.c
+++ b/kernels/bulldozer/3/bli_gemm_bulldozer_asm_d4x6_fma4.c
@@ -107,18 +107,17 @@ void bli_sgemm_bulldozer_asm_8x8_fma4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	
 	vmovaps(mem(rax, 0*32), ymm0) // initialize loop by pre-loading
 	vmovsldup(mem(rbx, 0*32), ymm2) // elements of a and b.
 	vpermilps(imm(0x4e), ymm2, ymm3)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 4), rdi) // cs_c *= sizeof(float)
 	lea(mem(rcx, rdi, 4), r10) // load address of c + 4*cs_c;
 	
@@ -142,7 +141,7 @@ void bli_sgemm_bulldozer_asm_8x8_fma4
 	vxorps(ymm15, ymm15, ymm15)
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -251,7 +250,7 @@ void bli_sgemm_bulldozer_asm_8x8_fma4
 	
 	label(.SCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -348,8 +347,8 @@ void bli_sgemm_bulldozer_asm_8x8_fma4
 	 //   ab61    ab63    ab65    ab67
 	 //   ab71 )  ab73 )  ab75 )  ab77 )
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rax), ymm0) // load alpha and duplicate
 	vbroadcastss(mem(rbx), ymm4) // load beta and duplicate
 	
@@ -365,7 +364,7 @@ void bli_sgemm_bulldozer_asm_8x8_fma4
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 4), rsi) // rsi = rs_c * sizeof(float)
 	
 	lea(mem(rcx, rsi, 4), rdx) // load address of c + 4*rs_c;
@@ -740,19 +739,20 @@ void bli_sgemm_bulldozer_asm_8x8_fma4
 	label(.SDONE)
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter]  "m" (k_iter), // 0
+      [k_left]  "m" (k_left), // 1
+      [a]       "m" (a),      // 2
+      [b]       "m" (b),      // 3
+      [alpha]   "m" (alpha),  // 4
+      [beta]    "m" (beta),   // 5
+      [c]       "m" (c),      // 6
+      [rs_c]    "m" (rs_c),   // 7
+      [cs_c]    "m" (cs_c)/*,   // 8
+      [b_next]  "m" (b_next), // 9
+      [a_next]  "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -761,7 +761,7 @@ void bli_sgemm_bulldozer_asm_8x8_fma4
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 #undef KERNEL4x6_1
@@ -879,13 +879,12 @@ void bli_dgemm_bulldozer_asm_4x6_fma4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ 
-	(
+	begin_asm()
 		
 		
 		vzeroall()
-		mov(%3, rbx) // load address of b.
-		mov(%2, rax) // load address of a.
+		mov(var(b), rbx) // load address of b.
+		mov(var(a), rax) // load address of a.
 		prefetch(0, mem(rax, 64))
 		
 		
@@ -895,7 +894,7 @@ void bli_dgemm_bulldozer_asm_4x6_fma4
 		add(imm(12*8), rbx)
 		add(imm(8*8), rax)
 		
-		mov(%0, rsi) // i = k_iter; notice %0 not $0
+		mov(var(k_iter), rsi) // i = k_iter; notice var(k_iter) not $0
 		test(rsi, rsi)
 		je(.CONSIDERKLEFT)
 		
@@ -920,7 +919,7 @@ void bli_dgemm_bulldozer_asm_4x6_fma4
 		
 		label(.CONSIDERKLEFT)
 		
-		mov(%1, rsi)
+		mov(var(k_left), rsi)
 		test(rsi, rsi) 
 		label(.LOOPKLEFT)
 		je(.POSTACCUM)
@@ -935,11 +934,11 @@ void bli_dgemm_bulldozer_asm_4x6_fma4
 		label(.POSTACCUM)
 		
 		
-		mov(%7, rsi) // load cs_c
-		mov(%8, rdi) // load rs_c
-		vmovddup(mem(%4), xmm2) //load alpha
-		vmovddup(mem(%5), xmm3) //load beta
-		mov(%6, rcx) // load address of c
+		mov(var(rs_c), rsi) // load cs_c
+		mov(var(cs_c), rdi) // load rs_c
+		vmovddup(mem(var(alpha)), xmm2) //load alpha
+		vmovddup(mem(var(beta)), xmm3) //load beta
+		mov(var(c), rcx) // load address of c
 		sal(imm(3), rsi) // cs_c *= sizeof(double)
 		sal(imm(3), rdi) // rs_c *= sizeof(double)
 		lea(mem(rcx, rdi, 2), rdx)
@@ -1034,17 +1033,20 @@ void bli_dgemm_bulldozer_asm_4x6_fma4
 		vmovhpd(xmm14, mem(rdx, rdi, 1))
 		vmovhpd(xmm15, mem(r8, rdi, 1)) 
 
+    end_asm(
 		: // output operands (none)
 		: // input operands
-		  "r" (k_iter),
-		  "r" (k_left),
-		  "r" (a),
-		  "r" (b),
-		  "r" (alpha),
-		  "r" (beta),
-		  "r" (c),
-		  "m" (rs_c),
-		  "m" (cs_c)
+	      [k_iter]  "r" (k_iter), // 0
+	      [k_left]  "r" (k_left), // 1
+	      [a]       "r" (a),      // 2
+	      [b]       "r" (b),      // 3
+	      [alpha]   "r" (alpha),  // 4
+	      [beta]    "r" (beta),   // 5
+	      [c]       "r" (c),      // 6
+	      [rs_c]    "m" (rs_c),   // 7
+	      [cs_c]    "m" (cs_c)/*,   // 8
+	      [b_next]  "m" (b_next), // 9
+	      [a_next]  "m" (a_next)*/  // 10
 		: // register clobber list
 		  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8",
 		  "xmm0", "xmm1", "xmm2", "xmm3",
@@ -1052,7 +1054,7 @@ void bli_dgemm_bulldozer_asm_4x6_fma4
 		  "xmm8", "xmm9", "xmm10", "xmm11",
 		  "xmm12", "xmm13", "xmm14", "xmm15",
 		  "memory"
-	);
+	)
 }
 //The parameter "i" is the iteration number, i.e. the B values to read
 #define MADD_TO_YMM(i) \
@@ -1094,21 +1096,20 @@ void bli_cgemm_bulldozer_asm_8x4_fma4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	mov(%9, r15) // load address of b_next.
-	//mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	mov(var(b_next), r15) // load address of b_next.
+	//mov(var(a_next), r14) // load address of a_next.
 	sub(imm(4*64), r15)
 	
 	vmovaps(mem(rax, 0*32), ymm0) // initialize loop by pre-loading
 	vmovsldup(mem(rbx, 0*32), ymm2)
 	vpermilps(imm(0x4e), ymm2, ymm3)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(scomplex)
 	lea(mem(rcx, rdi, 2), r10) // load address of c + 2*cs_c;
 	
@@ -1126,7 +1127,7 @@ void bli_cgemm_bulldozer_asm_8x4_fma4
 	vxorps(ymm14, ymm14, ymm14)
 	vxorps(ymm15, ymm15, ymm15)
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -1266,7 +1267,7 @@ void bli_cgemm_bulldozer_asm_8x4_fma4
 	
 	label(.CCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -1377,7 +1378,7 @@ void bli_cgemm_bulldozer_asm_8x4_fma4
 	
 	 // scale by alpha
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastss(mem(rax), ymm7) // load alpha_r and duplicate
 	vbroadcastss(mem(rax, 4), ymm6) // load alpha_i and duplicate
 	
@@ -1424,7 +1425,7 @@ void bli_cgemm_bulldozer_asm_8x4_fma4
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rbx), ymm7) // load beta_r and duplicate
 	vbroadcastss(mem(rbx, 4), ymm6) // load beta_i and duplicate
 	
@@ -1434,7 +1435,7 @@ void bli_cgemm_bulldozer_asm_8x4_fma4
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(scomplex)
 	
 	lea(mem(rcx, rsi, 4), rdx) // load address of c + 4*rs_c;
@@ -1835,19 +1836,20 @@ void bli_cgemm_bulldozer_asm_8x4_fma4
 	label(.CDONE)
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c),   // 8
-	  "m" (b_next)/*, // 9
-	  "m" (a_next)*/  // 10
+      [k_iter]  "m" (k_iter), // 0
+      [k_left]  "m" (k_left), // 1
+      [a]       "m" (a),      // 2
+      [b]       "m" (b),      // 3
+      [alpha]   "m" (alpha),  // 4
+      [beta]    "m" (beta),   // 5
+      [c]       "m" (c),      // 6
+      [rs_c]    "m" (rs_c),   // 7
+      [cs_c]    "m" (cs_c),   // 8
+      [b_next]  "m" (b_next)/*, // 9
+      [a_next]  "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1856,7 +1858,7 @@ void bli_cgemm_bulldozer_asm_8x4_fma4
 	  "ymm8", "ymm9", "ymm10", "ymm11",
 	  "ymm12", "ymm13", "ymm14", "ymm15",
 	  "memory"
-	);
+	)
 }
 
 #define MADDSUBPD_TO_YMM \
@@ -1905,21 +1907,20 @@ void bli_zgemm_bulldozer_asm_4x4_fma4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	//mov(%9, r15) // load address of b_next.
-	//mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	//mov(var(b_next), r15) // load address of b_next.
+	//mov(var(a_next), r14) // load address of a_next.
 	
 	vmovapd(mem(rax, 0*32), ymm0) // initialize loop by pre-loading
 	vmovddup(mem(rbx, 0+0*32), ymm2)
 	vmovddup(mem(rbx, 0+1*32), ymm3)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(dcomplex)
 	lea(mem(, rdi, 2), rdi)
 	lea(mem(rcx, rdi, 2), r10) // load address of c + 2*cs_c;
@@ -1939,7 +1940,7 @@ void bli_zgemm_bulldozer_asm_4x4_fma4
 	vxorpd(ymm15, ymm15, ymm15)
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -2083,7 +2084,7 @@ void bli_zgemm_bulldozer_asm_4x4_fma4
 	
 	label(.ZCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -2176,7 +2177,7 @@ void bli_zgemm_bulldozer_asm_4x4_fma4
 	
 	 // scale by alpha
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastsd(mem(rax), ymm7) // load alpha_r and duplicate
 	vbroadcastsd(mem(rax, 8), ymm6) // load alpha_i and duplicate
 	
@@ -2190,13 +2191,13 @@ void bli_zgemm_bulldozer_asm_4x4_fma4
 	Z_ALPHA(9, 1)
 	Z_ALPHA(8, 0)
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastsd(mem(rbx), ymm7) // load beta_r and duplicate
 	vbroadcastsd(mem(rbx, 8), ymm6) // load beta_i and duplicate
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(dcomplex)
 	lea(mem(, rsi, 2), rsi)
 	lea(mem(rcx, rsi, 2), rdx) // load address of c + 2*rs_c;
@@ -2508,19 +2509,20 @@ void bli_zgemm_bulldozer_asm_4x4_fma4
 	label(.ZDONE)
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+	  [k_iter]  "m" (k_iter), // 0
+	  [k_left]  "m" (k_left), // 1
+	  [a]       "m" (a),      // 2
+	  [b]       "m" (b),      // 3
+	  [alpha]   "m" (alpha),  // 4
+	  [beta]    "m" (beta),   // 5
+	  [c]       "m" (c),      // 6
+	  [rs_c]    "m" (rs_c),   // 7
+	  [cs_c]    "m" (cs_c)/*,   // 8
+	  [b_next]  "m" (b_next), // 9
+	  [a_next]  "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -2529,6 +2531,6 @@ void bli_zgemm_bulldozer_asm_4x4_fma4
 	  "ymm8", "ymm9", "ymm10", "ymm11",
 	  "ymm12", "ymm13", "ymm14", "ymm15",
 	  "memory"
-	);
+	)
 }
 
diff --git a/kernels/knl/1m/bli_dpackm_knl_asm_24x8.c b/kernels/knl/1m/bli_dpackm_knl_asm_24x8.c
index 29ee9cdc9..0d8996671 100644
--- a/kernels/knl/1m/bli_dpackm_knl_asm_24x8.c
+++ b/kernels/knl/1m/bli_dpackm_knl_asm_24x8.c
@@ -127,7 +127,7 @@ void bli_dpackm_knl_asm_8xk
     const int64_t lda = lda_;
     const int64_t ldp = ldp_;
 
-    BEGIN_ASM
+    BEGIN_ASM()
 
     MOV(RSI, VAR(n))
     MOV(RAX, VAR(a))
@@ -321,8 +321,8 @@ void bli_dpackm_knl_asm_24xk
     const int64_t lda = lda_;
     const int64_t ldp = ldp_;
 
-    __asm__ volatile
-    (
+    BEGIN_ASM()
+
         MOV(RSI, VAR(n))
         MOV(RAX, VAR(a))
         MOV(RBX, VAR(inca))
@@ -524,6 +524,7 @@ void bli_dpackm_knl_asm_24xk
 
         LABEL(PACK24_DONE)
 
+    END_ASM(
         : //output operands
         : //input operands
           [n]         "m" (n),
@@ -543,5 +544,5 @@ void bli_dpackm_knl_asm_24xk
           "zmm30", "zmm31",
           "rax", "rbx", "rcx", "rdi", "rsi",
           "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", "memory"
-    );
+    )
 }
diff --git a/kernels/knl/1m/bli_spackm_knl_asm_24x16.c b/kernels/knl/1m/bli_spackm_knl_asm_24x16.c
index 961df18ef..f9c69d008 100644
--- a/kernels/knl/1m/bli_spackm_knl_asm_24x16.c
+++ b/kernels/knl/1m/bli_spackm_knl_asm_24x16.c
@@ -129,8 +129,8 @@ void bli_spackm_knl_asm_16xk
     const int64_t lda = lda_;
     const int64_t ldp = ldp_;
 
-    __asm__ volatile
-    (
+    BEGIN_ASM()
+
         MOV(RSI, VAR(n))
         MOV(RAX, VAR(a))
         MOV(RBX, VAR(inca))
@@ -295,6 +295,7 @@ void bli_spackm_knl_asm_16xk
 
         LABEL(PACK16_DONE)
 
+    END_ASM(
         : //output operands
         : //input operands
           [n]         "m" (n),
@@ -314,7 +315,7 @@ void bli_spackm_knl_asm_16xk
           "zmm30", "zmm31",
           "rax", "rbx", "rcx", "rdx", "rdi", "rsi",
           "r8", "r9", "r10", "r11", "r12", "r13", "r14", "memory"
-    );
+    )
 }
 
 void bli_spackm_knl_asm_24xk
@@ -338,8 +339,8 @@ void bli_spackm_knl_asm_24xk
     const int64_t lda = lda_;
     const int64_t ldp = ldp_;
 
-    __asm__ volatile
-    (
+    BEGIN_ASM()
+
         MOV(RSI, VAR(n))
         MOV(RAX, VAR(a))
         MOV(RBX, VAR(inca))
@@ -540,6 +541,7 @@ void bli_spackm_knl_asm_24xk
 
         LABEL(PACK24_DONE)
 
+    END_ASM(
         : //output operands
         : //input operands
           [n]         "m" (n),
@@ -559,5 +561,5 @@ void bli_spackm_knl_asm_24xk
           "zmm30", "zmm31",
           "rax", "rbx", "rcx", "rdx", "rdi", "rsi",
           "r8", "r9", "r10", "r11", "r12", "r13", "r14", "memory"
-    );
+    )
 }
diff --git a/kernels/knl/3/bli_dgemm_knl_asm_24x8.c b/kernels/knl/3/bli_dgemm_knl_asm_24x8.c
index 196372c1a..a135b4f27 100644
--- a/kernels/knl/3/bli_dgemm_knl_asm_24x8.c
+++ b/kernels/knl/3/bli_dgemm_knl_asm_24x8.c
@@ -213,7 +213,7 @@ void bli_dgemm_knl_asm_24x8
     int tlooph, tloopl, blooph, bloopl;
 #endif
 
-    BEGIN_ASM
+    BEGIN_ASM()
 
 #ifdef MONITORS
     RDTSC
diff --git a/kernels/knl/3/bli_sgemm_knl_asm_24x16.c b/kernels/knl/3/bli_sgemm_knl_asm_24x16.c
index 41e4b12aa..620b436ee 100644
--- a/kernels/knl/3/bli_sgemm_knl_asm_24x16.c
+++ b/kernels/knl/3/bli_sgemm_knl_asm_24x16.c
@@ -210,8 +210,8 @@ void bli_sgemm_knl_asm_24x16
     int tlooph, tloopl, blooph, bloopl;
 #endif
 
-    __asm__ volatile
-    (
+    BEGIN_ASM()
+
 #ifdef MONITORS
     RDTSC
     MOV(VAR(topl), EAX)
@@ -664,6 +664,8 @@ void bli_sgemm_knl_asm_24x16
     MOV(VAR(botl), EAX)
     MOV(VAR(both), EDX)
 #endif
+
+    END_ASM(
     : // output operands
 #ifdef MONITORS
       [topl]  "=m" (topl),
@@ -694,7 +696,7 @@ void bli_sgemm_knl_asm_24x16
       "zmm14", "zmm15", "zmm16", "zmm17", "zmm18", "zmm19", "zmm20", "zmm21",
       "zmm22", "zmm23", "zmm24", "zmm25", "zmm26", "zmm27", "zmm28", "zmm29",
       "zmm30", "zmm31", "memory"
-    );
+    )
 
 #ifdef LOOPMON
     printf("looptime = \t%d\n", bloopl - tloopl);
diff --git a/kernels/penryn/3/bli_gemm_penryn_asm_d4x4.c b/kernels/penryn/3/bli_gemm_penryn_asm_d4x4.c
index 98486a2e8..84a998019 100644
--- a/kernels/penryn/3/bli_gemm_penryn_asm_d4x4.c
+++ b/kernels/penryn/3/bli_gemm_penryn_asm_d4x4.c
@@ -59,13 +59,12 @@ void bli_sgemm_penryn_asm_8x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 		
 		
-		mov(%2, rax) // load address of a.
-		mov(%3, rbx) // load address of b.
-		mov(%9, r9) // load address of b_next.
+		mov(var(a), rax) // load address of a.
+		mov(var(b), rbx) // load address of b.
+		mov(var(b_next), r9) // load address of b_next.
 		
 		sub(imm(0-8*16), rax) // increment pointers to allow byte
 		sub(imm(0-8*16), rbx) // offsets in the unrolled iterations.
@@ -74,8 +73,8 @@ void bli_sgemm_penryn_asm_8x4
 		movaps(mem(rax, -7*16), xmm1) // of a and b.
 		movaps(mem(rbx, -8*16), xmm2)
 		
-		mov(%6, rcx) // load address of c
-		mov(%8, rdi) // load cs_c
+		mov(var(c), rcx) // load address of c
+		mov(var(cs_c), rdi) // load cs_c
 		lea(mem(, rdi, 4), rdi) // cs_c *= sizeof(float)
 		mov(rdi, r12) // make a copy of cs_c (in bytes)
 		lea(mem(rcx, rdi, 2), r10) // load address of c + 2*cs_c;
@@ -102,7 +101,7 @@ void bli_sgemm_penryn_asm_8x4
 		
 		
 		
-		mov(%0, rsi) // i = k_iter;
+		mov(var(k_iter), rsi) // i = k_iter;
 		test(rsi, rsi) // check i via logical AND.
 		je(.SCONSIDKLEFT) // if i == 0, jump to code that
 		 // contains the k_left loop.
@@ -252,7 +251,7 @@ void bli_sgemm_penryn_asm_8x4
 		
 		label(.SCONSIDKLEFT)
 		
-		mov(%1, rsi) // i = k_left;
+		mov(var(k_left), rsi) // i = k_left;
 		test(rsi, rsi) // check i via logical AND.
 		je(.SPOSTACCUM) // if i == 0, we're done; jump to end.
 		 // else, we prepare to enter k_left loop.
@@ -307,15 +306,15 @@ void bli_sgemm_penryn_asm_8x4
 		addps(xmm5, xmm15)
 		
 		
-		mov(%4, rax) // load address of alpha
-		mov(%5, rbx) // load address of beta 
+		mov(var(alpha), rax) // load address of alpha
+		mov(var(beta), rbx) // load address of beta
 		movss(mem(rax), xmm6) // load alpha to bottom 4 bytes of xmm6
 		movss(mem(rbx), xmm7) // load beta to bottom 4 bytes of xmm7
 		pshufd(imm(0x00), xmm6, xmm6) // populate xmm6 with four alphas
 		pshufd(imm(0x00), xmm7, xmm7) // populate xmm7 with four betas
 		
 		
-		mov(%7, rsi) // load rs_c
+		mov(var(rs_c), rsi) // load rs_c
 		mov(rsi, r8) // make a copy of rs_c
 		
 		lea(mem(, rsi, 4), rsi) // rsi = rs_c * sizeof(float)
@@ -821,18 +820,20 @@ void bli_sgemm_penryn_asm_8x4
 		label(.SDONE)
 		
 
+    end_asm(
 		: // output operands (none)
 		: // input operands
-		  "m" (k_iter),
-		  "m" (k_left),
-		  "m" (a),
-		  "m" (b),
-		  "m" (alpha),
-		  "m" (beta),
-		  "m" (c),
-		  "m" (rs_c),
-		  "m" (cs_c),
-		  "m" (b_next)
+	      [k_iter] "m" (k_iter), // 0
+	      [k_left] "m" (k_left), // 1
+	      [a]      "m" (a),      // 2
+	      [b]      "m" (b),      // 3
+	      [alpha]  "m" (alpha),  // 4
+	      [beta]   "m" (beta),   // 5
+	      [c]      "m" (c),      // 6
+	      [rs_c]   "m" (rs_c),   // 7
+	      [cs_c]   "m" (cs_c),   // 8
+	      [b_next] "m" (b_next)/*, // 9
+	      [a_next] "m" (a_next)*/  // 10
 		: // register clobber list
 		  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11", "r12",
 		  "xmm0", "xmm1", "xmm2", "xmm3",
@@ -840,7 +841,7 @@ void bli_sgemm_penryn_asm_8x4
 		  "xmm8", "xmm9", "xmm10", "xmm11",
 		  "xmm12", "xmm13", "xmm14", "xmm15",
 		  "memory"
-	);
+	)
 }
 
 void bli_dgemm_penryn_asm_4x4
@@ -865,14 +866,13 @@ void bli_dgemm_penryn_asm_4x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 		
 		
-		mov(%2, rax) // load address of a.
-		mov(%3, rbx) // load address of b.
-		mov(%9, r9) // load address of b_next.
-		mov(%10, r11) // load address of a_next.
+		mov(var(a), rax) // load address of a.
+		mov(var(b), rbx) // load address of b.
+		mov(var(b_next), r9) // load address of b_next.
+		mov(var(a_next), r11) // load address of a_next.
 		
 		sub(imm(0-8*16), rax) // increment pointers to allow byte
 		sub(imm(0-8*16), rbx) // offsets in the unrolled iterations.
@@ -881,8 +881,8 @@ void bli_dgemm_penryn_asm_4x4
 		movaps(mem(rax, -7*16), xmm1) // of a and b.
 		movaps(mem(rbx, -8*16), xmm2)
 		
-		mov(%6, rcx) // load address of c
-		mov(%8, rdi) // load cs_c
+		mov(var(c), rcx) // load address of c
+		mov(var(cs_c), rdi) // load cs_c
 		lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(double)
 		mov(rdi, r12) // make a copy of cs_c (in bytes)
 		lea(mem(rcx, rdi, 2), r10) // load address of c + 2*cs_c;
@@ -909,7 +909,7 @@ void bli_dgemm_penryn_asm_4x4
 		
 		
 		
-		mov(%0, rsi) // i = k_iter;
+		mov(var(k_iter), rsi) // i = k_iter;
 		test(rsi, rsi) // check i via logical AND.
 		je(.DCONSIDKLEFT) // if i == 0, jump to code that
 		 // contains the k_left loop.
@@ -1076,7 +1076,7 @@ void bli_dgemm_penryn_asm_4x4
 		
 		label(.DCONSIDKLEFT)
 		
-		mov(%1, rsi) // i = k_left;
+		mov(var(k_left), rsi) // i = k_left;
 		test(rsi, rsi) // check i via logical AND.
 		je(.DPOSTACCUM) // if i == 0, we're done; jump to end.
 		 // else, we prepare to enter k_left loop.
@@ -1132,13 +1132,13 @@ void bli_dgemm_penryn_asm_4x4
 		addpd(xmm6, xmm14)
 		
 		
-		mov(%4, rax) // load address of alpha
-		mov(%5, rbx) // load address of beta 
+		mov(var(alpha), rax) // load address of alpha
+		mov(var(beta), rbx) // load address of beta
 		movddup(mem(rax), xmm6) // load alpha and duplicate
 		movddup(mem(rbx), xmm7) // load beta and duplicate
 		
 		
-		mov(%7, rsi) // load rs_c
+		mov(var(rs_c), rsi) // load rs_c
 		mov(rsi, r8) // make a copy of rs_c
 		
 		lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(double)
@@ -1467,19 +1467,20 @@ void bli_dgemm_penryn_asm_4x4
 		label(.DDONE)
 		
 
+    end_asm(
 		: // output operands (none)
 		: // input operands
-		  "m" (k_iter), // 0
-		  "m" (k_left), // 1
-		  "m" (a),      // 2
-		  "m" (b),      // 3
-		  "m" (alpha),  // 4
-		  "m" (beta),   // 5
-		  "m" (c),      // 6
-		  "m" (rs_c),   // 7
-		  "m" (cs_c),   // 8
-		  "m" (b_next), // 9
-		  "m" (a_next)  // 10
+	      [k_iter] "m" (k_iter), // 0
+	      [k_left] "m" (k_left), // 1
+	      [a]      "m" (a),      // 2
+	      [b]      "m" (b),      // 3
+	      [alpha]  "m" (alpha),  // 4
+	      [beta]   "m" (beta),   // 5
+	      [c]      "m" (c),      // 6
+	      [rs_c]   "m" (rs_c),   // 7
+	      [cs_c]   "m" (cs_c),   // 8
+	      [b_next] "m" (b_next), // 9
+	      [a_next] "m" (a_next)  // 10
 		: // register clobber list
 		  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11", "r12",
 		  "xmm0", "xmm1", "xmm2", "xmm3",
@@ -1487,7 +1488,7 @@ void bli_dgemm_penryn_asm_4x4
 		  "xmm8", "xmm9", "xmm10", "xmm11",
 		  "xmm12", "xmm13", "xmm14", "xmm15",
 		  "memory"
-	);
+	)
 }
 
 
diff --git a/kernels/penryn/3/bli_gemmtrsm_l_penryn_asm_d4x4.c b/kernels/penryn/3/bli_gemmtrsm_l_penryn_asm_d4x4.c
index 992d17967..ac8659396 100644
--- a/kernels/penryn/3/bli_gemmtrsm_l_penryn_asm_d4x4.c
+++ b/kernels/penryn/3/bli_gemmtrsm_l_penryn_asm_d4x4.c
@@ -76,12 +76,11 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 		
-		mov(%2, rax) // load address of a10.
-		mov(%4, rbx) // load address of b01.
-		//mov(%10, r9) // load address of b_next.
+		mov(var(a10), rax) // load address of a10.
+		mov(var(b01), rbx) // load address of b01.
+		//mov(var(b_next), r9) // load address of b_next.
 		
 		sub(imm(0-8*16), rax) // increment pointers to allow byte
 		sub(imm(0-8*16), rbx) // offsets in the unrolled iterations.
@@ -90,8 +89,8 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 		movaps(mem(rax, -7*16), xmm1) // of a and b.
 		movaps(mem(rbx, -8*16), xmm2)
 		
-		//mov(%6, rcx) // load address of c11
-		//mov(%9, rdi) // load cs_c
+		//mov(var(c11), rcx) // load address of c11
+		//mov(var(rs_c), rdi) // load cs_c
 		//lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(double)
 		//lea(mem(rcx, rdi, 2), rdx) // load address of c + 2*cs_c;
 		
@@ -117,7 +116,7 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 		
 		
 		
-		mov(%0, rsi) // i = k_iter;
+		mov(var(k_iter), rsi) // i = k_iter;
 		test(rsi, rsi) // check i via logical AND.
 		je(.CONSIDERKLEFT) // if i == 0, jump to code that
 		 // contains the k_left loop.
@@ -270,7 +269,7 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 		
 		label(.CONSIDERKLEFT)
 		
-		mov(%1, rsi) // i = k_left;
+		mov(var(k_left), rsi) // i = k_left;
 		test(rsi, rsi) // check i via logical AND.
 		je(.POSTACCUM) // if i == 0, we're done; jump to end.
 		 // else, we prepare to enter k_left loop.
@@ -327,7 +326,7 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 		
 		
 		
-		mov(%5, rbx) // load address of b11.
+		mov(var(b11), rbx) // load address of b11.
 		
 		 // xmm8:   xmm9:   xmm10:  xmm11:
 		 // ( ab01  ( ab00  ( ab03  ( ab02
@@ -361,7 +360,7 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 		 // xmm2: ( ab20 ab21 ) xmm6: ( ab22 ab23 )
 		 // xmm3: ( ab30 ab31 ) xmm7: ( ab32 ab33 )
 		
-		mov(%9, rax) // load address of alpha
+		mov(var(alpha), rax) // load address of alpha
 		movddup(mem(rax), xmm15) // load alpha and duplicate
 		
 		movaps(mem(rbx, 0*16), xmm8) 
@@ -400,11 +399,11 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 		label(.TRSM)
 		
 		
-		mov(%3, rax) // load address of a11
-		mov(%6, rcx) // load address of c11
+		mov(var(a11), rax) // load address of a11
+		mov(var(c11), rcx) // load address of c11
 		
-		mov(%7, rsi) // load rs_c
-		mov(%8, rdi) // load cs_c
+		mov(var(rs_c), rsi) // load rs_c
+		mov(var(cs_c), rdi) // load cs_c
 		sal(imm(3), rsi) // rs_c *= sizeof( double )
 		sal(imm(3), rdi) // cs_c *= sizeof( double )
 		
@@ -519,19 +518,20 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 		
 		
 
+    end_asm(
 		: // output operands (none)
 		: // input operands
-		  "m" (k_iter), // 0
-		  "m" (k_left), // 1
-		  "m" (a10),    // 2
-		  "m" (a11),    // 3
-		  "m" (b01),    // 4
-		  "m" (b11),    // 5
-		  "m" (c11),    // 6
-		  "m" (rs_c),   // 7
-		  "m" (cs_c),   // 8
-		  "m" (alpha),  // 9
-		  "m" (b_next)  // 10
+          [k_iter] "m" (k_iter), // 0
+          [k_left] "m" (k_left), // 1
+          [a10]    "m" (a10),    // 2
+          [a11]    "m" (a11),    // 3
+          [b01]    "m" (b01),    // 4
+          [b11]    "m" (b11),    // 5
+          [c11]    "m" (c11),    // 6
+          [rs_c]   "m" (rs_c),   // 7
+          [cs_c]   "m" (cs_c),   // 8
+          [alpha]  "m" (alpha),  // 9
+          [b_next] "m" (b_next)  // 10
 		: // register clobber list
 		  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", //"r8", "r9", "r10",
 		  "xmm0", "xmm1", "xmm2", "xmm3",
@@ -539,7 +539,7 @@ void bli_dgemmtrsm_l_penryn_asm_4x4
 		  "xmm8", "xmm9", "xmm10", "xmm11",
 		  "xmm12", "xmm13", "xmm14", "xmm15",
 		  "memory"
-	);
+	)
 
 }
 
diff --git a/kernels/penryn/3/bli_gemmtrsm_u_penryn_asm_d4x4.c b/kernels/penryn/3/bli_gemmtrsm_u_penryn_asm_d4x4.c
index c0b94269a..3d6e8a0d4 100644
--- a/kernels/penryn/3/bli_gemmtrsm_u_penryn_asm_d4x4.c
+++ b/kernels/penryn/3/bli_gemmtrsm_u_penryn_asm_d4x4.c
@@ -76,12 +76,11 @@ void bli_dgemmtrsm_u_penryn_asm_4x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 		
-		mov(%2, rax) // load address of a12.
-		mov(%4, rbx) // load address of b21.
-		//mov(%10, r9) // load address of b_next.
+		mov(var(a12), rax) // load address of a12.
+		mov(var(b21), rbx) // load address of b21.
+		//mov(var(b_next), r9) // load address of b_next.
 		
 		add(imm(8*16), rax) // increment pointers to allow byte
 		add(imm(8*16), rbx) // offsets in the unrolled iterations.
@@ -106,7 +105,7 @@ void bli_dgemmtrsm_u_penryn_asm_4x4
 		
 		
 		
-		mov(%0, rsi) // i = k_iter;
+		mov(var(k_iter), rsi) // i = k_iter;
 		test(rsi, rsi) // check i via logical AND.
 		je(.CONSIDERKLEFT) // if i == 0, jump to code that
 		 // contains the k_left loop.
@@ -253,7 +252,7 @@ void bli_dgemmtrsm_u_penryn_asm_4x4
 		
 		label(.CONSIDERKLEFT)
 		
-		mov(%1, rsi) // i = k_left;
+		mov(var(k_left), rsi) // i = k_left;
 		test(rsi, rsi) // check i via logical AND.
 		je(.POSTACCUM) // if i == 0, we're done; jump to end.
 		 // else, we prepare to enter k_left loop.
@@ -310,7 +309,7 @@ void bli_dgemmtrsm_u_penryn_asm_4x4
 		
 		
 		
-		mov(%5, rbx) // load address of b11.
+		mov(var(b11), rbx) // load address of b11.
 		
 		 // xmm8:   xmm9:   xmm10:  xmm11:
 		 // ( ab01  ( ab00  ( ab03  ( ab02
@@ -344,7 +343,7 @@ void bli_dgemmtrsm_u_penryn_asm_4x4
 		 // xmm2: ( ab20 ab21 ) xmm6: ( ab22 ab23 )
 		 // xmm3: ( ab30 ab31 ) xmm7: ( ab32 ab33 )
 		
-		mov(%9, rax) // load address of alpha
+		mov(var(alpha), rax) // load address of alpha
 		movddup(mem(rax), xmm15) // load alpha and duplicate
 		
 		movaps(mem(rbx, 0*16), xmm8)
@@ -383,11 +382,11 @@ void bli_dgemmtrsm_u_penryn_asm_4x4
 		label(.TRSM)
 		
 		
-		mov(%3, rax) // load address of a11
-		mov(%6, rcx) // load address of c11
+		mov(var(a11), rax) // load address of a11
+		mov(var(c11), rcx) // load address of c11
 		
-		mov(%7, rsi) // load rs_c
-		mov(%8, rdi) // load cs_c
+		mov(var(rs_c), rsi) // load rs_c
+		mov(var(cs_c), rdi) // load cs_c
 		sal(imm(3), rsi) // rs_c *= sizeof( double )
 		sal(imm(3), rdi) // cs_c *= sizeof( double )
 		
@@ -504,20 +503,20 @@ void bli_dgemmtrsm_u_penryn_asm_4x4
 		
 		
 		
-
+    end_asm(
 		: // output operands (none)
 		: // input operands
-		  "m" (k_iter), // 0
-		  "m" (k_left), // 1
-		  "m" (a12),    // 2
-		  "m" (a11),    // 3
-		  "m" (b21),    // 4
-		  "m" (b11),    // 5
-		  "m" (c11),    // 6
-		  "m" (rs_c),   // 7
-		  "m" (cs_c),   // 8
-		  "m" (alpha),  // 9
-		  "m" (b_next)  // 10
+		  [k_iter] "m" (k_iter), // 0
+		  [k_left] "m" (k_left), // 1
+		  [a12]    "m" (a12),    // 2
+		  [a11]    "m" (a11),    // 3
+		  [b21]    "m" (b21),    // 4
+		  [b11]    "m" (b11),    // 5
+		  [c11]    "m" (c11),    // 6
+		  [rs_c]   "m" (rs_c),   // 7
+		  [cs_c]   "m" (cs_c),   // 8
+		  [alpha]  "m" (alpha),  // 9
+		  [b_next] "m" (b_next)  // 10
 		: // register clobber list
 		  "rax", "rbx", "rcx", "rdx", "rsi", "rdi",
 		  "xmm0", "xmm1", "xmm2", "xmm3",
@@ -525,7 +524,7 @@ void bli_dgemmtrsm_u_penryn_asm_4x4
 		  "xmm8", "xmm9", "xmm10", "xmm11",
 		  "xmm12", "xmm13", "xmm14", "xmm15",
 		  "memory"
-	);
+	)
 
 }
 
diff --git a/kernels/penryn/3/bli_trsm_l_penryn_asm_d4x4.c b/kernels/penryn/3/bli_trsm_l_penryn_asm_d4x4.c
index ab8c846bb..39a06b538 100644
--- a/kernels/penryn/3/bli_trsm_l_penryn_asm_d4x4.c
+++ b/kernels/penryn/3/bli_trsm_l_penryn_asm_d4x4.c
@@ -64,10 +64,9 @@ void bli_dtrsm_l_penryn_asm_4x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 		
-		mov(%1, rbx) // load address of b11.
+		mov(var(b11), rbx) // load address of b11.
 		
 		movaps(mem(rbx, 0*16), xmm8) // xmm8  = ( beta00 beta01 )
 		movaps(mem(rbx, 1*16), xmm12) // xmm9  = ( beta02 beta03 )
@@ -80,11 +79,11 @@ void bli_dtrsm_l_penryn_asm_4x4
 		
 		
 		
-		mov(%0, rax) // load address of a11
-		mov(%2, rcx) // load address of c11
+		mov(var(a11), rax) // load address of a11
+		mov(var(c11), rcx) // load address of c11
 		
-		mov(%3, rsi) // load rs_c
-		mov(%4, rdi) // load cs_c
+		mov(var(rs_c), rsi) // load rs_c
+		mov(var(cs_c), rdi) // load cs_c
 		sal(imm(3), rsi) // rs_c *= sizeof( double )
 		sal(imm(3), rdi) // cs_c *= sizeof( double )
 		
@@ -199,13 +198,14 @@ void bli_dtrsm_l_penryn_asm_4x4
 		
 		
 
+    end_asm(
 		: // output operands (none)
 		: // input operands
-		  "m" (a11),    // 0
-		  "m" (b11),    // 1
-		  "m" (c11),    // 2
-		  "m" (rs_c),   // 3
-		  "m" (cs_c)    // 4
+          [a11]  "m" (a11),    // 0
+          [b11]  "m" (b11),    // 1
+          [c11]  "m" (c11),    // 2
+          [rs_c] "m" (rs_c),   // 3
+          [cs_c] "m" (cs_c)    // 4
 		: // register clobber list
 		  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", //"r8", "r9", "r10",
 		  "xmm0", "xmm1", "xmm2", "xmm3",
@@ -213,7 +213,7 @@ void bli_dtrsm_l_penryn_asm_4x4
 		  "xmm8", "xmm9", "xmm10", "xmm11",
 		  "xmm12", "xmm13", "xmm14", "xmm15",
 		  "memory"
-	);
+	)
 
 }
 
diff --git a/kernels/penryn/3/bli_trsm_u_penryn_asm_d4x4.c b/kernels/penryn/3/bli_trsm_u_penryn_asm_d4x4.c
index de78c59db..290aa27e0 100644
--- a/kernels/penryn/3/bli_trsm_u_penryn_asm_d4x4.c
+++ b/kernels/penryn/3/bli_trsm_u_penryn_asm_d4x4.c
@@ -64,10 +64,9 @@ void bli_dtrsm_u_penryn_asm_4x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 		
-		mov(%1, rbx) // load address of b11.
+		mov(var(b11), rbx) // load address of b11.
 		
 		movaps(mem(rbx, 0*16), xmm8) // xmm8  = ( beta00 beta01 )
 		movaps(mem(rbx, 1*16), xmm12) // xmm9  = ( beta02 beta03 )
@@ -80,11 +79,11 @@ void bli_dtrsm_u_penryn_asm_4x4
 		
 		
 		
-		mov(%0, rax) // load address of a11
-		mov(%2, rcx) // load address of c11
+		mov(var(a11), rax) // load address of a11
+		mov(var(c11), rcx) // load address of c11
 		
-		mov(%3, rsi) // load rs_c
-		mov(%4, rdi) // load cs_c
+		mov(var(rs_c), rsi) // load rs_c
+		mov(var(cs_c), rdi) // load cs_c
 		sal(imm(3), rsi) // rs_c *= sizeof( double )
 		sal(imm(3), rdi) // cs_c *= sizeof( double )
 		
@@ -202,13 +201,14 @@ void bli_dtrsm_u_penryn_asm_4x4
 		
 		
 
+    end_asm(
 		: // output operands (none)
 		: // input operands
-		  "m" (a11),    // 0
-		  "m" (b11),    // 1
-		  "m" (c11),    // 2
-		  "m" (rs_c),   // 3
-		  "m" (cs_c)    // 4
+		  [a11]  "m" (a11),    // 0
+		  [b11]  "m" (b11),    // 1
+		  [c11]  "m" (c11),    // 2
+		  [rs_c] "m" (rs_c),   // 3
+		  [cs_c] "m" (cs_c)    // 4
 		: // register clobber list
 		  "rax", "rbx", "rcx", "rdx", "rsi", "rdi",
 		  "xmm0", "xmm1", "xmm2", "xmm3",
@@ -216,7 +216,7 @@ void bli_dtrsm_u_penryn_asm_4x4
 		  "xmm8", "xmm9", "xmm10", "xmm11",
 		  "xmm12", "xmm13", "xmm14", "xmm15",
 		  "memory"
-	);
+	)
 
 }
 
diff --git a/kernels/piledriver/3/bli_gemm_piledriver_asm_d8x3.c b/kernels/piledriver/3/bli_gemm_piledriver_asm_d8x3.c
index 82866f1fd..4174c1527 100644
--- a/kernels/piledriver/3/bli_gemm_piledriver_asm_d8x3.c
+++ b/kernels/piledriver/3/bli_gemm_piledriver_asm_d8x3.c
@@ -62,14 +62,13 @@ void bli_sgemm_piledriver_asm_16x3
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	mov(%9, r15) // load address of b_next.
-	mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	mov(var(b_next), r15) // load address of b_next.
+	mov(var(a_next), r14) // load address of a_next.
 	
 	prefetch(0, mem(rbx, 128)) // prefetch b
 	prefetch(0, mem(rbx, 64+128)) // prefetch b
@@ -78,8 +77,8 @@ void bli_sgemm_piledriver_asm_16x3
 	add(imm(32*4), rax)
 	add(imm(12*4), rbx)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 4), rdi) // cs_c *= sizeof(float)
 	lea(mem(rcx, rdi, 1), r10) // load address of c + 1*cs_c;
 	lea(mem(rcx, rdi, 2), r11) // load address of c + 2*cs_c;
@@ -103,7 +102,7 @@ void bli_sgemm_piledriver_asm_16x3
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -313,7 +312,7 @@ void bli_sgemm_piledriver_asm_16x3
 	
 	label(.SCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -393,8 +392,8 @@ void bli_sgemm_piledriver_asm_16x3
 	
 	
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rax), xmm0) // load alpha and duplicate
 	vbroadcastss(mem(rbx), xmm2) // load beta and duplicate
 	
@@ -419,7 +418,7 @@ void bli_sgemm_piledriver_asm_16x3
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 4), rsi) // rsi = rs_c * sizeof(float)
 	
 	//lea(mem(rcx, rsi, 4), rdx) // load address of c + 4*rs_c;
@@ -887,19 +886,20 @@ void bli_sgemm_piledriver_asm_16x3
 	label(.SDONE)
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c),   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c),   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -908,7 +908,7 @@ void bli_sgemm_piledriver_asm_16x3
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 void bli_dgemm_piledriver_asm_8x3
@@ -933,14 +933,13 @@ void bli_dgemm_piledriver_asm_8x3
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	mov(%9, r15) // load address of b_next.
-	mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	mov(var(b_next), r15) // load address of b_next.
+	mov(var(a_next), r14) // load address of a_next.
 	
 	prefetch(0, mem(rbx, 128)) // prefetch b
 	prefetch(0, mem(rbx, 64+128)) // prefetch b
@@ -949,8 +948,8 @@ void bli_dgemm_piledriver_asm_8x3
 	add(imm(16*8), rax)
 	add(imm(12*8), rbx)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(double)
 	lea(mem(rcx, rdi, 1), r10) // load address of c + 1*cs_c;
 	lea(mem(rcx, rdi, 2), r11) // load address of c + 2*cs_c;
@@ -974,7 +973,7 @@ void bli_dgemm_piledriver_asm_8x3
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -1183,7 +1182,7 @@ void bli_dgemm_piledriver_asm_8x3
 	
 	label(.DCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DPOSTACCUM) // if i == 0, we're done.
 	 // else, we prepare to
@@ -1253,8 +1252,8 @@ void bli_dgemm_piledriver_asm_8x3
 	
 	
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vmovddup(mem(rax), xmm0) // load alpha and duplicate
 	vmovddup(mem(rbx), xmm2) // load beta and duplicate
 	
@@ -1278,7 +1277,7 @@ void bli_dgemm_piledriver_asm_8x3
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(double)
 	
 	lea(mem(rcx, rsi, 4), rdx) // load address of c + 4*rs_c;
@@ -1606,19 +1605,20 @@ void bli_dgemm_piledriver_asm_8x3
 	label(.DDONE)
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c),   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c),   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1627,7 +1627,7 @@ void bli_dgemm_piledriver_asm_8x3
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 void bli_cgemm_piledriver_asm_4x2
@@ -1652,17 +1652,16 @@ void bli_cgemm_piledriver_asm_4x2
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	mov(%9, r15) // load address of b_next.
-	mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	mov(var(b_next), r15) // load address of b_next.
+	mov(var(a_next), r14) // load address of a_next.
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(scomplex)
 	lea(mem(rcx, rdi, 1), r10) // load address of c + 1*cs_c;
 	
@@ -1682,7 +1681,7 @@ void bli_cgemm_piledriver_asm_4x2
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -1848,7 +1847,7 @@ void bli_cgemm_piledriver_asm_4x2
 	
 	label(.CCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -1925,7 +1924,7 @@ void bli_cgemm_piledriver_asm_4x2
 	
 	 // scale by alpha
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastss(mem(rax), xmm0) // load alpha_r and duplicate
 	vbroadcastss(mem(rax, 4), xmm1) // load alpha_i and duplicate
 	
@@ -1952,7 +1951,7 @@ void bli_cgemm_piledriver_asm_4x2
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rbx), xmm6) // load beta_r and duplicate
 	vbroadcastss(mem(rbx, 4), xmm7) // load beta_i and duplicate
 	
@@ -1962,7 +1961,7 @@ void bli_cgemm_piledriver_asm_4x2
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(scomplex)
 	
 	
@@ -2154,19 +2153,20 @@ void bli_cgemm_piledriver_asm_4x2
 	label(.CDONE)
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c),   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c),   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -2175,7 +2175,7 @@ void bli_cgemm_piledriver_asm_4x2
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 void bli_zgemm_piledriver_asm_2x2
@@ -2200,17 +2200,16 @@ void bli_zgemm_piledriver_asm_2x2
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	mov(%9, r15) // load address of b_next.
-	mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	mov(var(b_next), r15) // load address of b_next.
+	mov(var(a_next), r14) // load address of a_next.
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(dcomplex)
 	lea(mem(, rdi, 2), rdi)
 	lea(mem(rcx, rdi, 1), r10) // load address of c + 1*cs_c;
@@ -2229,7 +2228,7 @@ void bli_zgemm_piledriver_asm_2x2
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -2399,7 +2398,7 @@ void bli_zgemm_piledriver_asm_2x2
 	
 	label(.ZCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -2473,7 +2472,7 @@ void bli_zgemm_piledriver_asm_2x2
 	
 	 // scale by alpha
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vmovddup(mem(rax), xmm0) // load alpha_r and duplicate
 	vmovddup(mem(rax, 8), xmm1) // load alpha_i and duplicate
 	
@@ -2500,7 +2499,7 @@ void bli_zgemm_piledriver_asm_2x2
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vmovddup(mem(rbx), xmm6) // load beta_r and duplicate
 	vmovddup(mem(rbx, 8), xmm7) // load beta_i and duplicate
 	
@@ -2510,7 +2509,7 @@ void bli_zgemm_piledriver_asm_2x2
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(dcomplex)
 	lea(mem(, rsi, 2), rsi)
 	//lea(mem(rcx, rsi, 2), rdx) // load address of c + 2*rs_c;
@@ -2688,19 +2687,20 @@ void bli_zgemm_piledriver_asm_2x2
 	label(.ZDONE)
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c),   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c),   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -2709,7 +2709,7 @@ void bli_zgemm_piledriver_asm_2x2
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
diff --git a/kernels/sandybridge/3/bli_gemm_sandybridge_asm_d8x4.c b/kernels/sandybridge/3/bli_gemm_sandybridge_asm_d8x4.c
index 20946c3c5..1532664d7 100644
--- a/kernels/sandybridge/3/bli_gemm_sandybridge_asm_d8x4.c
+++ b/kernels/sandybridge/3/bli_gemm_sandybridge_asm_d8x4.c
@@ -62,20 +62,19 @@ void bli_sgemm_sandybridge_asm_8x8
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	//mov(%9, r15) // load address of b_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	//mov(var(b_next), r15) // load address of b_next.
 	
 	vmovaps(mem(rax, 0*32), ymm0) // initialize loop by pre-loading
 	vmovsldup(mem(rbx, 0*32), ymm2) // elements of a and b.
 	vpermilps(imm(0x4e), ymm2, ymm3)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 4), rdi) // cs_c *= sizeof(float)
 	lea(mem(rcx, rdi, 4), r10) // load address of c + 4*cs_c;
 	
@@ -100,7 +99,7 @@ void bli_sgemm_sandybridge_asm_8x8
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -248,7 +247,7 @@ void bli_sgemm_sandybridge_asm_8x8
 	
 	label(.SCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -393,8 +392,8 @@ void bli_sgemm_sandybridge_asm_8x8
 	
 	
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rax), ymm0) // load alpha and duplicate
 	vbroadcastss(mem(rbx), ymm4) // load beta and duplicate
 	
@@ -412,7 +411,7 @@ void bli_sgemm_sandybridge_asm_8x8
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 4), rsi) // rsi = rs_c * sizeof(float)
 	
 	lea(mem(rcx, rsi, 4), rdx) // load address of c + 4*rs_c;
@@ -1002,19 +1001,20 @@ void bli_sgemm_sandybridge_asm_8x8
 	vzeroupper()
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1023,7 +1023,7 @@ void bli_sgemm_sandybridge_asm_8x8
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 void bli_dgemm_sandybridge_asm_8x4
@@ -1048,22 +1048,21 @@ void bli_dgemm_sandybridge_asm_8x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	mov(%9, r15) // load address of b_next.
-	//mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	mov(var(b_next), r15) // load address of b_next.
+	//mov(var(a_next), r14) // load address of a_next.
 	sub(imm(4*64), r15)
 	
 	vmovapd(mem(rax, 0*32), ymm0) // initialize loop by pre-loading
 	vmovapd(mem(rbx, 0*32), ymm2) // elements of a and b.
 	vpermilpd(imm(0x5), ymm2, ymm3)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(double)
 	lea(mem(rcx, rdi, 2), r10) // load address of c + 2*cs_c;
 	
@@ -1083,7 +1082,7 @@ void bli_dgemm_sandybridge_asm_8x4
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -1228,7 +1227,7 @@ void bli_dgemm_sandybridge_asm_8x4
 	
 	label(.DCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -1343,8 +1342,8 @@ void bli_dgemm_sandybridge_asm_8x4
 	 //   ab70 )  ab71 )  ab72 )  ab73 )
 	
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastsd(mem(rax), ymm0) // load alpha and duplicate
 	vbroadcastsd(mem(rbx), ymm2) // load beta and duplicate
 	
@@ -1362,7 +1361,7 @@ void bli_dgemm_sandybridge_asm_8x4
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(double)
 	
 	lea(mem(rcx, rsi, 4), rdx) // load address of c + 4*rs_c;
@@ -1677,19 +1676,20 @@ void bli_dgemm_sandybridge_asm_8x4
 	vzeroupper()
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c),   // 8
-	  "m" (b_next)/*, // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c),   // 8
+      [b_next] "m" (b_next)/*, // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1698,7 +1698,7 @@ void bli_dgemm_sandybridge_asm_8x4
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 void bli_cgemm_sandybridge_asm_8x4
@@ -1723,22 +1723,21 @@ void bli_cgemm_sandybridge_asm_8x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	mov(%9, r15) // load address of b_next.
-	//mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	mov(var(b_next), r15) // load address of b_next.
+	//mov(var(a_next), r14) // load address of a_next.
 	sub(imm(4*64), r15)
 	
 	vmovaps(mem(rax, 0*32), ymm0) // initialize loop by pre-loading
 	vmovsldup(mem(rbx, 0*32), ymm2)
 	vpermilps(imm(0x4e), ymm2, ymm3)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(scomplex)
 	lea(mem(rcx, rdi, 2), r10) // load address of c + 2*cs_c;
 	
@@ -1758,7 +1757,7 @@ void bli_cgemm_sandybridge_asm_8x4
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -2004,7 +2003,7 @@ void bli_cgemm_sandybridge_asm_8x4
 	
 	label(.CCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -2175,7 +2174,7 @@ void bli_cgemm_sandybridge_asm_8x4
 	
 	 // scale by alpha
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastss(mem(rax), ymm7) // load alpha_r and duplicate
 	vbroadcastss(mem(rax, 4), ymm6) // load alpha_i and duplicate
 	
@@ -2222,7 +2221,7 @@ void bli_cgemm_sandybridge_asm_8x4
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rbx), ymm7) // load beta_r and duplicate
 	vbroadcastss(mem(rbx, 4), ymm6) // load beta_i and duplicate
 	
@@ -2232,7 +2231,7 @@ void bli_cgemm_sandybridge_asm_8x4
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(scomplex)
 	
 	lea(mem(rcx, rsi, 4), rdx) // load address of c + 4*rs_c;
@@ -2638,19 +2637,20 @@ void bli_cgemm_sandybridge_asm_8x4
 	vzeroupper()
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c),   // 8
-	  "m" (b_next)/*, // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c),   // 8
+      [b_next] "m" (b_next)/*, // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -2659,7 +2659,7 @@ void bli_cgemm_sandybridge_asm_8x4
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
@@ -2686,21 +2686,20 @@ void bli_zgemm_sandybridge_asm_4x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
-	//mov(%9, r15) // load address of b_next.
-	//mov(%10, r14) // load address of a_next.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
+	//mov(var(b_next), r15) // load address of b_next.
+	//mov(var(a_next), r14) // load address of a_next.
 	
 	vmovapd(mem(rax, 0*32), ymm0) // initialize loop by pre-loading
 	vmovddup(mem(rbx, 0+0*32), ymm2)
 	vmovddup(mem(rbx, 0+1*32), ymm3)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(dcomplex)
 	lea(mem(, rdi, 2), rdi)
 	lea(mem(rcx, rdi, 2), r10) // load address of c + 2*cs_c;
@@ -2721,7 +2720,7 @@ void bli_zgemm_sandybridge_asm_4x4
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -2964,7 +2963,7 @@ void bli_zgemm_sandybridge_asm_4x4
 	
 	label(.ZCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -3083,7 +3082,7 @@ void bli_zgemm_sandybridge_asm_4x4
 	
 	 // scale by alpha
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastsd(mem(rax), ymm7) // load alpha_r and duplicate
 	vbroadcastsd(mem(rax, 8), ymm6) // load alpha_i and duplicate
 	
@@ -3130,7 +3129,7 @@ void bli_zgemm_sandybridge_asm_4x4
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastsd(mem(rbx), ymm7) // load beta_r and duplicate
 	vbroadcastsd(mem(rbx, 8), ymm6) // load beta_i and duplicate
 	
@@ -3140,7 +3139,7 @@ void bli_zgemm_sandybridge_asm_4x4
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(dcomplex)
 	lea(mem(, rsi, 2), rsi)
 	lea(mem(rcx, rsi, 2), rdx) // load address of c + 2*rs_c;
@@ -3488,19 +3487,20 @@ void bli_zgemm_sandybridge_asm_4x4
 	vzeroupper()
 	
 
+	end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -3509,7 +3509,7 @@ void bli_zgemm_sandybridge_asm_4x4
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
diff --git a/kernels/skx/3/bli_dgemm_skx_asm_16x12_l2.c b/kernels/skx/3/bli_dgemm_skx_asm_16x12_l2.c
index 54bdb9e1d..d841dad80 100644
--- a/kernels/skx/3/bli_dgemm_skx_asm_16x12_l2.c
+++ b/kernels/skx/3/bli_dgemm_skx_asm_16x12_l2.c
@@ -306,7 +306,7 @@ void bli_dgemm_skx_asm_16x12_l2(
     const int64_t rs_c = rs_c_;
     const int64_t cs_c = cs_c_;
 
-    BEGIN_ASM
+    BEGIN_ASM()
 
     VXORPD(YMM(8), YMM(8), YMM(8)) //clear out registers
     VMOVAPD(YMM( 7), YMM(8))
diff --git a/kernels/skx/3/bli_sgemm_skx_asm_32x12_l2.c b/kernels/skx/3/bli_sgemm_skx_asm_32x12_l2.c
index a879ada00..39ed53b27 100644
--- a/kernels/skx/3/bli_sgemm_skx_asm_32x12_l2.c
+++ b/kernels/skx/3/bli_sgemm_skx_asm_32x12_l2.c
@@ -336,7 +336,7 @@ void bli_sgemm_skx_asm_32x12_l2(
     const int64_t rs_c = rs_c_;
     const int64_t cs_c = cs_c_;
 
-    BEGIN_ASM
+    BEGIN_ASM()
 
     VXORPD(YMM(8), YMM(8), YMM(8)) //clear out registers
     VMOVAPD(YMM( 7), YMM(8))
diff --git a/kernels/zen/3/bli_gemm_zen_asm_d6x8.c b/kernels/zen/3/bli_gemm_zen_asm_d6x8.c
index c06ff6f90..f21620263 100644
--- a/kernels/zen/3/bli_gemm_zen_asm_d6x8.c
+++ b/kernels/zen/3/bli_gemm_zen_asm_d6x8.c
@@ -99,14 +99,13 @@ void bli_sgemm_zen_asm_6x16
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	//mov(%9, r15) // load address of b_next.
 	
 	add(imm(32*4), rbx)
@@ -114,8 +113,8 @@ void bli_sgemm_zen_asm_6x16
 	vmovaps(mem(rbx, -4*32), ymm0)
 	vmovaps(mem(rbx, -3*32), ymm1)
 	
-	mov(%6, rcx) // load address of c
-	mov(%7, rdi) // load rs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(rs_c), rdi) // load rs_c
 	lea(mem(, rdi, 4), rdi) // rs_c *= sizeof(float)
 	
 	lea(mem(rdi, rdi, 2), r13) // r13 = 3*rs_c;
@@ -130,7 +129,7 @@ void bli_sgemm_zen_asm_6x16
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -257,7 +256,7 @@ void bli_sgemm_zen_asm_6x16
 	
 	label(.SCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -305,8 +304,8 @@ void bli_sgemm_zen_asm_6x16
 	
 	
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rax), ymm0) // load alpha and duplicate
 	vbroadcastss(mem(rbx), ymm3) // load beta and duplicate
 	
@@ -328,7 +327,7 @@ void bli_sgemm_zen_asm_6x16
 	
 	
 	
-	mov(%8, rsi) // load cs_c
+	mov(var(cs_c), rsi) // load cs_c
 	lea(mem(, rsi, 4), rsi) // rsi = cs_c * sizeof(float)
 	
 	lea(mem(rcx, rsi, 8), rdx) // load address of c +  8*cs_c;
@@ -872,19 +871,20 @@ void bli_sgemm_zen_asm_6x16
 	
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -893,7 +893,7 @@ void bli_sgemm_zen_asm_6x16
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
@@ -945,14 +945,13 @@ void bli_dgemm_zen_asm_6x8
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	//mov(%9, r15) // load address of b_next.
 	
 	add(imm(32*4), rbx)
@@ -960,8 +959,8 @@ void bli_dgemm_zen_asm_6x8
 	vmovapd(mem(rbx, -4*32), ymm0)
 	vmovapd(mem(rbx, -3*32), ymm1)
 	
-	mov(%6, rcx) // load address of c
-	mov(%7, rdi) // load rs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(rs_c), rdi) // load rs_c
 	lea(mem(, rdi, 8), rdi) // rs_c *= sizeof(double)
 	
 	lea(mem(rdi, rdi, 2), r13) // r13 = 3*rs_c;
@@ -976,7 +975,7 @@ void bli_dgemm_zen_asm_6x8
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -1103,7 +1102,7 @@ void bli_dgemm_zen_asm_6x8
 	
 	label(.DCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -1151,8 +1150,8 @@ void bli_dgemm_zen_asm_6x8
 	
 	
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastsd(mem(rax), ymm0) // load alpha and duplicate
 	vbroadcastsd(mem(rbx), ymm3) // load beta and duplicate
 	
@@ -1174,7 +1173,7 @@ void bli_dgemm_zen_asm_6x8
 	
 	
 	
-	mov(%8, rsi) // load cs_c
+	mov(var(cs_c), rsi) // load cs_c
 	lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(double)
 	
 	lea(mem(rcx, rsi, 4), rdx) // load address of c +  4*cs_c;
@@ -1611,19 +1610,20 @@ void bli_dgemm_zen_asm_6x8
 	
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1632,7 +1632,7 @@ void bli_dgemm_zen_asm_6x8
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
@@ -1691,14 +1691,13 @@ void bli_cgemm_zen_asm_3x8
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	//mov(%9, r15) // load address of b_next.
 	
 	add(imm(32*4), rbx)
@@ -1706,8 +1705,8 @@ void bli_cgemm_zen_asm_3x8
 	vmovaps(mem(rbx, -4*32), ymm0)
 	vmovaps(mem(rbx, -3*32), ymm1)
 	
-	mov(%6, rcx) // load address of c
-	mov(%7, rdi) // load rs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(rs_c), rdi) // load rs_c
 	lea(mem(, rdi, 8), rdi) // rs_c *= sizeof(scomplex)
 	
 	lea(mem(rcx, rdi, 1), r11) // r11 = c + 1*rs_c;
@@ -1720,7 +1719,7 @@ void bli_cgemm_zen_asm_3x8
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -1847,7 +1846,7 @@ void bli_cgemm_zen_asm_3x8
 	
 	label(.CCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -1916,7 +1915,7 @@ void bli_cgemm_zen_asm_3x8
 	
 	
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastss(mem(rax), ymm0) // load alpha_r and duplicate
 	vbroadcastss(mem(rax, 4), ymm1) // load alpha_i and duplicate
 	
@@ -1957,14 +1956,14 @@ void bli_cgemm_zen_asm_3x8
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rbx), ymm1) // load beta_r and duplicate
 	vbroadcastss(mem(rbx, 4), ymm2) // load beta_i and duplicate
 	
 	
 	
 	
-	mov(%8, rsi) // load cs_c
+	mov(var(cs_c), rsi) // load cs_c
 	lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(scomplex)
 	lea(mem(, rsi, 4), rdx) // rdx = 4*cs_c;
 	lea(mem(rsi, rsi, 2), r13) // r13 = 3*cs_c;
@@ -2143,19 +2142,20 @@ void bli_cgemm_zen_asm_3x8
 	
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -2164,7 +2164,7 @@ void bli_cgemm_zen_asm_3x8
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
@@ -2219,14 +2219,13 @@ void bli_zgemm_zen_asm_3x4
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	//mov(%9, r15) // load address of b_next.
 	
 	add(imm(32*4), rbx)
@@ -2234,8 +2233,8 @@ void bli_zgemm_zen_asm_3x4
 	vmovapd(mem(rbx, -4*32), ymm0)
 	vmovapd(mem(rbx, -3*32), ymm1)
 	
-	mov(%6, rcx) // load address of c
-	mov(%7, rdi) // load rs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(rs_c), rdi) // load rs_c
 	lea(mem(, rdi, 8), rdi) // rs_c *= sizeof(dcomplex)
 	lea(mem(, rdi, 2), rdi)
 	
@@ -2249,7 +2248,7 @@ void bli_zgemm_zen_asm_3x4
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -2376,7 +2375,7 @@ void bli_zgemm_zen_asm_3x4
 	
 	label(.ZCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -2444,7 +2443,7 @@ void bli_zgemm_zen_asm_3x4
 	
 	
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastsd(mem(rax), ymm0) // load alpha_r and duplicate
 	vbroadcastsd(mem(rax, 8), ymm1) // load alpha_i and duplicate
 	
@@ -2485,14 +2484,14 @@ void bli_zgemm_zen_asm_3x4
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastsd(mem(rbx), ymm1) // load beta_r and duplicate
 	vbroadcastsd(mem(rbx, 8), ymm2) // load beta_i and duplicate
 	
 	
 	
 	
-	mov(%8, rsi) // load cs_c
+	mov(var(cs_c), rsi) // load cs_c
 	lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(dcomplex)
 	lea(mem(, rsi, 2), rsi)
 	lea(mem(, rsi, 2), rdx) // rdx = 2*cs_c;
@@ -2671,19 +2670,20 @@ void bli_zgemm_zen_asm_3x4
 	
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -2692,7 +2692,7 @@ void bli_zgemm_zen_asm_3x4
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
diff --git a/kernels/zen/3/bli_gemm_zen_asm_d8x6.c b/kernels/zen/3/bli_gemm_zen_asm_d8x6.c
index 46179b1af..20a764671 100644
--- a/kernels/zen/3/bli_gemm_zen_asm_d8x6.c
+++ b/kernels/zen/3/bli_gemm_zen_asm_d8x6.c
@@ -98,14 +98,13 @@ void bli_sgemm_zen_asm_16x6
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	//mov(%9, r15) // load address of b_next.
 	
 	add(imm(32*4), rax)
@@ -113,8 +112,8 @@ void bli_sgemm_zen_asm_16x6
 	vmovaps(mem(rax, -4*32), ymm0)
 	vmovaps(mem(rax, -3*32), ymm1)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 4), rdi) // cs_c *= sizeof(float)
 	
 	lea(mem(rdi, rdi, 2), r13) // r13 = 3*cs_c;
@@ -129,7 +128,7 @@ void bli_sgemm_zen_asm_16x6
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -256,7 +255,7 @@ void bli_sgemm_zen_asm_16x6
 	
 	label(.SCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -304,8 +303,8 @@ void bli_sgemm_zen_asm_16x6
 	
 	
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rax), ymm0) // load alpha and duplicate
 	vbroadcastss(mem(rbx), ymm3) // load beta and duplicate
 	
@@ -327,7 +326,7 @@ void bli_sgemm_zen_asm_16x6
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 4), rsi) // rsi = rs_c * sizeof(float)
 	
 	lea(mem(rcx, rsi, 8), rdx) // load address of c +  8*rs_c;
@@ -614,19 +613,20 @@ void bli_sgemm_zen_asm_16x6
 	
 	
 
+	end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -635,7 +635,7 @@ void bli_sgemm_zen_asm_16x6
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 #define DGEMM_INPUT_GS_BETA_NZ \
@@ -684,14 +684,13 @@ void bli_dgemm_zen_asm_8x6
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	//mov(%9, r15) // load address of b_next.
 	
 	add(imm(32*4), rax)
@@ -699,8 +698,8 @@ void bli_dgemm_zen_asm_8x6
 	vmovapd(mem(rax, -4*32), ymm0)
 	vmovapd(mem(rax, -3*32), ymm1)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(double)
 	
 	lea(mem(rdi, rdi, 2), r13) // r13 = 3*cs_c;
@@ -715,7 +714,7 @@ void bli_dgemm_zen_asm_8x6
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -842,7 +841,7 @@ void bli_dgemm_zen_asm_8x6
 	
 	label(.DCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -890,8 +889,8 @@ void bli_dgemm_zen_asm_8x6
 	
 	
 	
-	mov(%4, rax) // load address of alpha
-	mov(%5, rbx) // load address of beta 
+	mov(var(alpha), rax) // load address of alpha
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastsd(mem(rax), ymm0) // load alpha and duplicate
 	vbroadcastsd(mem(rbx), ymm3) // load beta and duplicate
 	
@@ -913,7 +912,7 @@ void bli_dgemm_zen_asm_8x6
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(double)
 	
 	lea(mem(rcx, rsi, 4), rdx) // load address of c +  4*rs_c;
@@ -1199,19 +1198,20 @@ void bli_dgemm_zen_asm_8x6
 	
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1220,7 +1220,7 @@ void bli_dgemm_zen_asm_8x6
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
@@ -1279,14 +1279,13 @@ void bli_cgemm_zen_asm_8x3
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	//mov(%9, r15) // load address of b_next.
 	
 	add(imm(32*4), rax)
@@ -1294,8 +1293,8 @@ void bli_cgemm_zen_asm_8x3
 	vmovaps(mem(rax, -4*32), ymm0)
 	vmovaps(mem(rax, -3*32), ymm1)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(scomplex)
 	
 	lea(mem(rcx, rdi, 1), r11) // r11 = c + 1*cs_c;
@@ -1308,7 +1307,7 @@ void bli_cgemm_zen_asm_8x3
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -1435,7 +1434,7 @@ void bli_cgemm_zen_asm_8x3
 	
 	label(.CCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.CPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -1504,7 +1503,7 @@ void bli_cgemm_zen_asm_8x3
 	
 	
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastss(mem(rax), ymm0) // load alpha_r and duplicate
 	vbroadcastss(mem(rax, 4), ymm1) // load alpha_i and duplicate
 	
@@ -1545,14 +1544,14 @@ void bli_cgemm_zen_asm_8x3
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastss(mem(rbx), ymm1) // load beta_r and duplicate
 	vbroadcastss(mem(rbx, 4), ymm2) // load beta_i and duplicate
 	
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(scomplex)
 	lea(mem(, rsi, 4), rdx) // rdx = 4*rs_c;
 	lea(mem(rsi, rsi, 2), r13) // r13 = 3*rs_c;
@@ -1731,19 +1730,20 @@ void bli_cgemm_zen_asm_8x3
 	
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a]      "m" (a),      // 2
+      [b]      "m" (b),      // 3
+      [alpha]  "m" (alpha),  // 4
+      [beta]   "m" (beta),   // 5
+      [c]      "m" (c),      // 6
+      [rs_c]   "m" (rs_c),   // 7
+      [cs_c]   "m" (cs_c)/*,   // 8
+      [b_next] "m" (b_next), // 9
+      [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1752,7 +1752,7 @@ void bli_cgemm_zen_asm_8x3
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
@@ -1807,14 +1807,13 @@ void bli_zgemm_zen_asm_4x3
 	uint64_t rs_c   = rs_c0;
 	uint64_t cs_c   = cs_c0;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a), rax) // load address of a.
+	mov(var(b), rbx) // load address of b.
 	//mov(%9, r15) // load address of b_next.
 	
 	add(imm(32*4), rax)
@@ -1822,8 +1821,8 @@ void bli_zgemm_zen_asm_4x3
 	vmovapd(mem(rax, -4*32), ymm0)
 	vmovapd(mem(rax, -3*32), ymm1)
 	
-	mov(%6, rcx) // load address of c
-	mov(%8, rdi) // load cs_c
+	mov(var(c), rcx) // load address of c
+	mov(var(cs_c), rdi) // load cs_c
 	lea(mem(, rdi, 8), rdi) // cs_c *= sizeof(dcomplex)
 	lea(mem(, rdi, 2), rdi)
 	
@@ -1837,7 +1836,7 @@ void bli_zgemm_zen_asm_4x3
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -1964,7 +1963,7 @@ void bli_zgemm_zen_asm_4x3
 	
 	label(.ZCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.ZPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -2032,7 +2031,7 @@ void bli_zgemm_zen_asm_4x3
 	
 	
 	
-	mov(%4, rax) // load address of alpha
+	mov(var(alpha), rax) // load address of alpha
 	vbroadcastsd(mem(rax), ymm0) // load alpha_r and duplicate
 	vbroadcastsd(mem(rax, 8), ymm1) // load alpha_i and duplicate
 	
@@ -2073,14 +2072,14 @@ void bli_zgemm_zen_asm_4x3
 	
 	
 	
-	mov(%5, rbx) // load address of beta 
+	mov(var(beta), rbx) // load address of beta
 	vbroadcastsd(mem(rbx), ymm1) // load beta_r and duplicate
 	vbroadcastsd(mem(rbx, 8), ymm2) // load beta_i and duplicate
 	
 	
 	
 	
-	mov(%7, rsi) // load rs_c
+	mov(var(rs_c), rsi) // load rs_c
 	lea(mem(, rsi, 8), rsi) // rsi = rs_c * sizeof(dcomplex)
 	lea(mem(, rsi, 2), rsi)
 	lea(mem(, rsi, 2), rdx) // rdx = 2*rs_c;
@@ -2259,19 +2258,20 @@ void bli_zgemm_zen_asm_4x3
 	
 	
 
+    end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a),      // 2
-	  "m" (b),      // 3
-	  "m" (alpha),  // 4
-	  "m" (beta),   // 5
-	  "m" (c),      // 6
-	  "m" (rs_c),   // 7
-	  "m" (cs_c)/*,   // 8
-	  "m" (b_next), // 9
-	  "m" (a_next)*/  // 10
+	  [k_iter] "m" (k_iter), // 0
+	  [k_left] "m" (k_left), // 1
+	  [a]      "m" (a),      // 2
+	  [b]      "m" (b),      // 3
+	  [alpha]  "m" (alpha),  // 4
+	  [beta]   "m" (beta),   // 5
+	  [c]      "m" (c),      // 6
+	  [rs_c]   "m" (rs_c),   // 7
+	  [cs_c]   "m" (cs_c)/*,   // 8
+	  [b_next] "m" (b_next), // 9
+	  [a_next] "m" (a_next)*/  // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -2280,7 +2280,7 @@ void bli_zgemm_zen_asm_4x3
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
diff --git a/kernels/zen/3/bli_gemmtrsm_l_zen_asm_d6x8.c b/kernels/zen/3/bli_gemmtrsm_l_zen_asm_d6x8.c
index f8717384c..2a3c8653a 100644
--- a/kernels/zen/3/bli_gemmtrsm_l_zen_asm_d6x8.c
+++ b/kernels/zen/3/bli_gemmtrsm_l_zen_asm_d6x8.c
@@ -81,36 +81,35 @@ void bli_sgemmtrsm_l_zen_asm_6x16
 
 	float*   beta   = bli_sm1;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a10), rax) // load address of a.
+	mov(var(b01), rbx) // load address of b.
 	
 	add(imm(32*4), rbx)
 	 // initialize loop by pre-loading
 	vmovaps(mem(rbx, -4*32), ymm0)
 	vmovaps(mem(rbx, -3*32), ymm1)
 	
-	mov(%7, rcx) // load address of b11
+	mov(var(b11), rcx) // load address of b11
 	mov(imm(16), rdi) // set rs_b = PACKNR = 16
 	lea(mem(, rdi, 4), rdi) // rs_b *= sizeof(float)
 	
 	 // NOTE: c11, rs_c, and cs_c aren't
 	 // needed for a while, but we load
 	 // them now to avoid stalling later.
-	mov(%8, r8) // load address of c11
-	mov(%9, r9) // load rs_c
+	mov(var(c11), r8) // load address of c11
+	mov(var(rs_c), r9) // load rs_c
 	lea(mem(, r9 , 4), r9) // rs_c *= sizeof(float)
-	mov(%10, r10) // load cs_c
+	mov(var(k_left)0, r10) // load cs_c
 	lea(mem(, r10, 4), r10) // cs_c *= sizeof(float)
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -237,7 +236,7 @@ void bli_sgemmtrsm_l_zen_asm_6x16
 	
 	label(.SCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -286,7 +285,7 @@ void bli_sgemmtrsm_l_zen_asm_6x16
 	
 	
 	
-	mov(%5, rbx) // load address of alpha
+	mov(var(alpha), rbx) // load address of alpha
 	vbroadcastss(mem(rbx), ymm3) // load alpha and duplicate
 	
 	
@@ -365,7 +364,7 @@ void bli_sgemmtrsm_l_zen_asm_6x16
 	 // ymm14 ymm15 = ( beta50..57 ) ( beta58..5F )
 	
 	
-	mov(%6, rax) // load address of a11
+	mov(var(a11), rax) // load address of a11
 	
 	mov(r11, rcx) // recall address of b11
 	mov(r14, rdx) // recall address of b11+8*cs_b
@@ -772,19 +771,20 @@ void bli_sgemmtrsm_l_zen_asm_6x16
 	vzeroupper()
 	
 
+	end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a10),    // 2
-	  "m" (b01),    // 3
-	  "m" (beta),   // 4
-	  "m" (alpha),  // 5
-	  "m" (a11),    // 6
-	  "m" (b11),    // 7
-	  "m" (c11),    // 8
-	  "m" (rs_c),   // 9
-	  "m" (cs_c)    // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a10]    "m" (a10),    // 2
+      [b01]    "m" (b01),    // 3
+      [beta]   "m" (beta),   // 4
+      [alpha]  "m" (alpha),  // 5
+      [a11]    "m" (a11),    // 6
+      [b11]    "m" (b11),    // 7
+      [c11]    "m" (c11),    // 8
+      [rs_c]   "m" (rs_c),   // 9
+      [cs_c]   "m" (cs_c)    // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -793,7 +793,7 @@ void bli_sgemmtrsm_l_zen_asm_6x16
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
@@ -835,36 +835,35 @@ void bli_dgemmtrsm_l_zen_asm_6x8
 
 	double*  beta   = bli_dm1;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a10), rax) // load address of a.
+	mov(var(b01), rbx) // load address of b.
 	
 	add(imm(32*4), rbx)
 	 // initialize loop by pre-loading
 	vmovapd(mem(rbx, -4*32), ymm0)
 	vmovapd(mem(rbx, -3*32), ymm1)
 	
-	mov(%7, rcx) // load address of b11
+	mov(var(b11), rcx) // load address of b11
 	mov(imm(8), rdi) // set rs_b = PACKNR = 8
 	lea(mem(, rdi, 8), rdi) // rs_b *= sizeof(double)
 	
 	 // NOTE: c11, rs_c, and cs_c aren't
 	 // needed for a while, but we load
 	 // them now to avoid stalling later.
-	mov(%8, r8) // load address of c11
-	mov(%9, r9) // load rs_c
+	mov(var(c11), r8) // load address of c11
+	mov(var(rs_c), r9) // load rs_c
 	lea(mem(, r9 , 8), r9) // rs_c *= sizeof(double)
-	mov(%10, r10) // load cs_c
+	mov(var(k_left)0, r10) // load cs_c
 	lea(mem(, r10, 8), r10) // cs_c *= sizeof(double)
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -991,7 +990,7 @@ void bli_dgemmtrsm_l_zen_asm_6x8
 	
 	label(.DCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -1041,7 +1040,7 @@ void bli_dgemmtrsm_l_zen_asm_6x8
 	
 	
 	
-	mov(%5, rbx) // load address of alpha
+	mov(var(alpha), rbx) // load address of alpha
 	vbroadcastsd(mem(rbx), ymm3) // load alpha and duplicate
 	
 	
@@ -1120,7 +1119,7 @@ void bli_dgemmtrsm_l_zen_asm_6x8
 	 // ymm14 ymm15 = ( beta50..53 ) ( beta54..57 )
 	
 	
-	mov(%6, rax) // load address of a11
+	mov(var(a11), rax) // load address of a11
 	
 	mov(r11, rcx) // recall address of b11
 	mov(r14, rdx) // recall address of b11+4*cs_b
@@ -1488,19 +1487,20 @@ void bli_dgemmtrsm_l_zen_asm_6x8
 	
 
 
+	end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a10),    // 2
-	  "m" (b01),    // 3
-	  "m" (beta),   // 4
-	  "m" (alpha),  // 5
-	  "m" (a11),    // 6
-	  "m" (b11),    // 7
-	  "m" (c11),    // 8
-	  "m" (rs_c),   // 9
-	  "m" (cs_c)    // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a10]    "m" (a10),    // 2
+      [b01]    "m" (b01),    // 3
+      [beta]   "m" (beta),   // 4
+      [alpha]  "m" (alpha),  // 5
+      [a11]    "m" (a11),    // 6
+      [b11]    "m" (b11),    // 7
+      [c11]    "m" (c11),    // 8
+      [rs_c]   "m" (rs_c),   // 9
+      [cs_c]   "m" (cs_c)    // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi",
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1509,7 +1509,7 @@ void bli_dgemmtrsm_l_zen_asm_6x8
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
diff --git a/kernels/zen/3/bli_gemmtrsm_u_zen_asm_d6x8.c b/kernels/zen/3/bli_gemmtrsm_u_zen_asm_d6x8.c
index 2cc742214..c4e84fc3a 100644
--- a/kernels/zen/3/bli_gemmtrsm_u_zen_asm_d6x8.c
+++ b/kernels/zen/3/bli_gemmtrsm_u_zen_asm_d6x8.c
@@ -81,36 +81,35 @@ void bli_sgemmtrsm_u_zen_asm_6x16
 
 	float*   beta   = bli_sm1;
 
-	__asm__ volatile
-	(
+	begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a10), rax) // load address of a.
+	mov(var(b01), rbx) // load address of b.
 	
 	add(imm(32*4), rbx)
 	 // initialize loop by pre-loading
 	vmovaps(mem(rbx, -4*32), ymm0)
 	vmovaps(mem(rbx, -3*32), ymm1)
 	
-	mov(%7, rcx) // load address of b11
+	mov(var(b11), rcx) // load address of b11
 	mov(imm(16), rdi) // set rs_b = PACKNR = 16
 	lea(mem(, rdi, 4), rdi) // rs_b *= sizeof(float)
 	
 	 // NOTE: c11, rs_c, and cs_c aren't
 	 // needed for a while, but we load
 	 // them now to avoid stalling later.
-	mov(%8, r8) // load address of c11
-	mov(%9, r9) // load rs_c
+	mov(var(c11), r8) // load address of c11
+	mov(var(rs_c), r9) // load rs_c
 	lea(mem(, r9 , 4), r9) // rs_c *= sizeof(float)
-	mov(%10, r10) // load cs_c
+	mov(var(k_left)0, r10) // load cs_c
 	lea(mem(, r10, 4), r10) // cs_c *= sizeof(float)
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -237,7 +236,7 @@ void bli_sgemmtrsm_u_zen_asm_6x16
 	
 	label(.SCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.SPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -286,7 +285,7 @@ void bli_sgemmtrsm_u_zen_asm_6x16
 	
 	
 	
-	mov(%5, rbx) // load address of alpha
+	mov(var(alpha), rbx) // load address of alpha
 	vbroadcastss(mem(rbx), ymm3) // load alpha and duplicate
 	
 	
@@ -365,7 +364,7 @@ void bli_sgemmtrsm_u_zen_asm_6x16
 	 // ymm14 ymm15 = ( beta50..57 ) ( beta58..5F )
 	
 	
-	mov(%6, rax) // load address of a11
+	mov(var(a11), rax) // load address of a11
 	
 	mov(r11, rcx) // recall address of b11
 	mov(r14, rdx) // recall address of b11+8*cs_b
@@ -776,20 +775,20 @@ void bli_sgemmtrsm_u_zen_asm_6x16
 	
 	vzeroupper()
 	
-
+	end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a10),    // 2
-	  "m" (b01),    // 3
-	  "m" (beta),   // 4
-	  "m" (alpha),  // 5
-	  "m" (a11),    // 6
-	  "m" (b11),    // 7
-	  "m" (c11),    // 8
-	  "m" (rs_c),   // 9
-	  "m" (cs_c)    // 10
+      [k_iter] "m" (k_iter), // 0
+      [k_left] "m" (k_left), // 1
+      [a10]    "m" (a10),    // 2
+      [b01]    "m" (b01),    // 3
+      [beta]   "m" (beta),   // 4
+      [alpha]  "m" (alpha),  // 5
+      [a11]    "m" (a11),    // 6
+      [b11]    "m" (b11),    // 7
+      [c11]    "m" (c11),    // 8
+      [rs_c]   "m" (rs_c),   // 9
+      [cs_c]   "m" (cs_c)    // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi", 
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -798,7 +797,7 @@ void bli_sgemmtrsm_u_zen_asm_6x16
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }
 
 
@@ -840,36 +839,35 @@ void bli_dgemmtrsm_u_zen_asm_6x8
 
 	double*  beta   = bli_dm1;
 
-	__asm__ volatile
-	(
+    begin_asm()
 	
 	vzeroall() // zero all xmm/ymm registers.
 	
 	
-	mov(%2, rax) // load address of a.
-	mov(%3, rbx) // load address of b.
+	mov(var(a10), rax) // load address of a.
+	mov(var(b01), rbx) // load address of b.
 	
 	add(imm(32*4), rbx)
 	 // initialize loop by pre-loading
 	vmovapd(mem(rbx, -4*32), ymm0)
 	vmovapd(mem(rbx, -3*32), ymm1)
 	
-	mov(%7, rcx) // load address of b11
+	mov(var(b11), rcx) // load address of b11
 	mov(imm(8), rdi) // set rs_b = PACKNR = 8
 	lea(mem(, rdi, 8), rdi) // rs_b *= sizeof(double)
 	
 	 // NOTE: c11, rs_c, and cs_c aren't
 	 // needed for a while, but we load
 	 // them now to avoid stalling later.
-	mov(%8, r8) // load address of c11
-	mov(%9, r9) // load rs_c
+	mov(var(c11), r8) // load address of c11
+	mov(var(rs_c), r9) // load rs_c
 	lea(mem(, r9 , 8), r9) // rs_c *= sizeof(double)
-	mov(%10, r10) // load cs_c
+	mov(var(k_left)0, r10) // load cs_c
 	lea(mem(, r10, 8), r10) // cs_c *= sizeof(double)
 	
 	
 	
-	mov(%0, rsi) // i = k_iter;
+	mov(var(k_iter), rsi) // i = k_iter;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DCONSIDKLEFT) // if i == 0, jump to code that
 	 // contains the k_left loop.
@@ -996,7 +994,7 @@ void bli_dgemmtrsm_u_zen_asm_6x8
 	
 	label(.DCONSIDKLEFT)
 	
-	mov(%1, rsi) // i = k_left;
+	mov(var(k_left), rsi) // i = k_left;
 	test(rsi, rsi) // check i via logical AND.
 	je(.DPOSTACCUM) // if i == 0, we're done; jump to end.
 	 // else, we prepare to enter k_left loop.
@@ -1046,7 +1044,7 @@ void bli_dgemmtrsm_u_zen_asm_6x8
 	
 	
 	
-	mov(%5, rbx) // load address of alpha
+	mov(var(alpha), rbx) // load address of alpha
 	vbroadcastsd(mem(rbx), ymm3) // load alpha and duplicate
 	
 	
@@ -1125,7 +1123,7 @@ void bli_dgemmtrsm_u_zen_asm_6x8
 	 // ymm14 ymm15 = ( beta50..53 ) ( beta54..57 )
 	
 	
-	mov(%6, rax) // load address of a11
+	mov(var(a11), rax) // load address of a11
 	
 	mov(r11, rcx) // recall address of b11
 	mov(r14, rdx) // recall address of b11+4*cs_b
@@ -1497,20 +1495,20 @@ void bli_dgemmtrsm_u_zen_asm_6x8
 	vzeroupper()
 	
 
-
+	end_asm(
 	: // output operands (none)
 	: // input operands
-	  "m" (k_iter), // 0
-	  "m" (k_left), // 1
-	  "m" (a10),    // 2
-	  "m" (b01),    // 3
-	  "m" (beta),   // 4
-	  "m" (alpha),  // 5
-	  "m" (a11),    // 6
-	  "m" (b11),    // 7
-	  "m" (c11),    // 8
-	  "m" (rs_c),   // 9
-	  "m" (cs_c)    // 10
+	  [k_iter] "m" (k_iter), // 0
+	  [k_left] "m" (k_left), // 1
+	  [a10]    "m" (a10),    // 2
+	  [b01]    "m" (b01),    // 3
+	  [beta]   "m" (beta),   // 4
+	  [alpha]  "m" (alpha),  // 5
+	  [a11]    "m" (a11),    // 6
+	  [b11]    "m" (b11),    // 7
+	  [c11]    "m" (c11),    // 8
+	  [rs_c]   "m" (rs_c),   // 9
+	  [cs_c]   "m" (cs_c)    // 10
 	: // register clobber list
 	  "rax", "rbx", "rcx", "rdx", "rsi", "rdi",
 	  "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
@@ -1519,7 +1517,7 @@ void bli_dgemmtrsm_u_zen_asm_6x8
 	  "xmm8", "xmm9", "xmm10", "xmm11",
 	  "xmm12", "xmm13", "xmm14", "xmm15",
 	  "memory"
-	);
+	)
 }