Fix fortran calling conv

EnzymeAD · Sep 20, 2023 · 210d033 · 210d033
1 parent c2924c5
commit 210d033
Show file tree

Hide file tree

Showing 16 changed files with 95 additions and 82 deletions.
diff --git a/enzyme/test/Enzyme/ReverseMode/blas/gemm_f.ll b/enzyme/test/Enzyme/ReverseMode/blas/gemm_f.ll
@@ -1,7 +1,7 @@
 ;RUN: if [ %llvmver -lt 16 ]; then %opt < %s %loadEnzyme -enzyme -S | FileCheck %s; fi
 ;RUN: %opt < %s %newLoadEnzyme -passes="enzyme" -S | FileCheck %s
 
-declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly) 
+declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i64, i64) 
 
 define void @f(i8* %C, i8* %A, i8* %B) {
 entry:
@@ -33,7 +33,7 @@ entry:
   store i64 8, i64* %ldb, align 16
   store double 0.000000e+00, double* %beta
   store i64 4, i64* %ldc, align 16
-  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p) 
+  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p, i64 1, i64 1) 
   ret void
 }
 
@@ -84,7 +84,7 @@ entry:
 ; CHECK-NEXT:   store i64 8, i64* %ldb, align 16
 ; CHECK-NEXT:   store double 0.000000e+00, double* %beta
 ; CHECK-NEXT:   store i64 4, i64* %ldc, align 16
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p, i64 1, i64 1)
 ; CHECK-NEXT:   br label %invertentry
 
 ; CHECK: invertentry:                                      ; preds = %entry
@@ -110,8 +110,8 @@ entry:
 ; CHECK-NEXT:   store i8 %[[i25]], i8* %byref.transpose.transb
 ; CHECK-NEXT:   store i64 1, i64* %byref.int.one
 ; CHECK-NEXT:   %intcast.int.one = bitcast i64* %byref.int.one to i8*
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %byref.transpose.transb, i8* %m_p, i8* %k_p, i8* %n_p, i8* %alpha_p, i8* %"C'", i8* %ldc_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %"A'", i8* %lda_p)
-; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %transb, i8* %k_p, i8* %n_p, i8* %m_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %"C'", i8* %ldc_p, i8* %beta_p, i8* %"B'", i8* %ldb_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %byref.transpose.transb, i8* %m_p, i8* %k_p, i8* %n_p, i8* %alpha_p, i8* %"C'", i8* %ldc_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %"A'", i8* %lda_p, i64 1, i64 1)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %transb, i8* %k_p, i8* %n_p, i8* %m_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %"C'", i8* %ldc_p, i8* %beta_p, i8* %"B'", i8* %ldb_p, i64 1, i64 1)
 ; CHECK-NEXT:   store i8 71, i8* %byref.constant.char.G
 ; CHECK-NEXT:   store i64 0, i64* %byref.constant.int.0
 ; CHECK-NEXT:   %intcast.constant.int.0 = bitcast i64* %byref.constant.int.0 to i8*

diff --git a/enzyme/test/Enzyme/ReverseMode/blas/gemm_f_c.ll b/enzyme/test/Enzyme/ReverseMode/blas/gemm_f_c.ll
@@ -1,7 +1,7 @@
 ;RUN: if [ %llvmver -lt 16 ]; then %opt < %s %loadEnzyme -enzyme -S | FileCheck %s; fi
 ;RUN: %opt < %s %newLoadEnzyme -passes="enzyme" -S | FileCheck %s
 
-declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly) 
+declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i64, i64) 
 
 define void @f(i8* %C, i8* %A, i8* %B) {
 entry:
@@ -33,7 +33,7 @@ entry:
   store i64 8, i64* %ldb, align 16
   store double 0.000000e+00, double* %beta
   store i64 4, i64* %ldc, align 16
-  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p) 
+  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p, i64 1, i64 1) 
   %ptr = bitcast i8* %A to double*
   store double 0.0000000e+00, double* %ptr, align 8
   ret void
@@ -176,7 +176,7 @@ entry:
 ; CHECK-NEXT:   br i1 %39, label %[[enzyme_memcpy_double_mat_64_exit21]], label %[[init_idx]]
 
 ; CHECK: [[enzyme_memcpy_double_mat_64_exit21]]:             ; preds = %__enzyme_memcpy_double_mat_64.exit, %[[init_end_i18]]
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p, i64 1, i64 1)
 ; CHECK-NEXT:   %"ptr'ipc" = bitcast i8* %"A'" to double*
 ; CHECK-NEXT:   %ptr = bitcast i8* %A to double*
 ; CHECK-NEXT:   store double 0.000000e+00, double* %ptr, align 8, !alias.scope !10, !noalias !13
@@ -213,13 +213,13 @@ entry:
 ; CHECK-DAG:   %[[r19:.+]] = icmp eq i8 %loaded.trans4, 110
 ; CHECK-DAG:   %[[r20:.+]] = or i1 %[[r19]], %[[r18]]
 ; CHECK-DAG:   %[[r21:.+]] = select i1 %[[r20]], i8* %k_p, i8* %n_p
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %byref.transpose.transb, i8* %m_p, i8* %k_p, i8* %n_p, i8* %alpha_p, i8* %"C'", i8* %ldc_p, i8* %[[i43]], i8* %[[r21]], i8* %beta_p, i8* %"A'", i8* %lda_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %byref.transpose.transb, i8* %m_p, i8* %k_p, i8* %n_p, i8* %alpha_p, i8* %"C'", i8* %ldc_p, i8* %[[i43]], i8* %[[r21]], i8* %beta_p, i8* %"A'", i8* %lda_p, i64 1, i64 1)
 ; CHECK-NEXT:   %loaded.trans5 = load i8, i8* %transa
 ; CHECK-DAG:   %[[r22:.+]] = icmp eq i8 %loaded.trans5, 78
 ; CHECK-DAG:   %[[r23:.+]] = icmp eq i8 %loaded.trans5, 110
 ; CHECK-DAG:   %[[r24:.+]] = or i1 %[[r23]], %[[r22]]
 ; CHECK-DAG:   %[[r25:.+]] = select i1 %[[r24]], i8* %m_p, i8* %k_p
-; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %transb, i8* %k_p, i8* %n_p, i8* %m_p, i8* %alpha_p, i8* %[[i42]], i8* %[[r25]], i8* %"C'", i8* %ldc_p, i8* %beta_p, i8* %"B'", i8* %ldb_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %transb, i8* %k_p, i8* %n_p, i8* %m_p, i8* %alpha_p, i8* %[[i42]], i8* %[[r25]], i8* %"C'", i8* %ldc_p, i8* %beta_p, i8* %"B'", i8* %ldb_p, i64 1, i64 1)
 ; CHECK-NEXT:   store i8 71, i8* %byref.constant.char.G
 ; CHECK-NEXT:   store i64 0, i64* %byref.constant.int.0
 ; CHECK-NEXT:   %intcast.constant.int.0 = bitcast i64* %byref.constant.int.0 to i8*

diff --git a/enzyme/test/Enzyme/ReverseMode/blas/gemm_f_c_lacpy.ll b/enzyme/test/Enzyme/ReverseMode/blas/gemm_f_c_lacpy.ll
@@ -1,7 +1,7 @@
 ;RUN: if [ %llvmver -lt 16 ]; then %opt < %s %loadEnzyme -enzyme -enzyme-lapack-copy=1 -S | FileCheck %s; fi
 ;RUN: %opt < %s %newLoadEnzyme -passes="enzyme" -enzyme-lapack-copy=1 -S | FileCheck %s
 
-declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly) 
+declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i64, i64) 
 
 define void @f(i8* noalias %C, i8* noalias %A, i8* noalias %B, i8* noalias %alpha, i8* noalias %beta) {
 entry:
@@ -27,7 +27,7 @@ entry:
   store i64 4, i64* %lda, align 16
   store i64 8, i64* %ldb, align 16
   store i64 4, i64* %ldc, align 16
-  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %C, i8* %ldc_p) 
+  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %C, i8* %ldc_p, i64 1, i64 1) 
   %ptr = bitcast i8* %A to double*
   store double 0.0000000e+00, double* %ptr, align 8
   ret void
@@ -117,7 +117,7 @@ entry:
 ; CHECK-NEXT:   %malloccall6 = tail call noalias nonnull i8* @malloc(i64 %mallocsize5)
 ; CHECK-NEXT:   %mat_AB = bitcast i8* %malloccall6 to double*
 ; CHECK-NEXT:   %[[i21:.+]] = bitcast double* %mat_AB to i8*
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %C, i8* %ldc_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %C, i8* %ldc_p, i64 1, i64 1)
 ; CHECK-NEXT:   %"ptr'ipc" = bitcast i8* %"A'" to double*
 ; CHECK-NEXT:   %ptr = bitcast i8* %A to double*
 ; CHECK-NEXT:   store double 0.000000e+00, double* %ptr, align 8, !alias.scope !0, !noalias !3
@@ -159,7 +159,7 @@ entry:
 ; CHECK-NEXT:   %[[i44:.+]] = select i1 %[[i43]], i8* %m_p, i8* %k_p
 ; CHECK-NEXT:   store double 0.000000e+00, double* %byref.constant.fp.0.0
 ; CHECK-NEXT:   %fpcast.constant.fp.0.0 = bitcast double* %byref.constant.fp.0.0 to i8*
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %fpcast.constant.fp.1.0, i8* %[[matA]], i8* %[[i44]], i8* %B, i8* %ldb_p, i8* %fpcast.constant.fp.0.0, i8* %[[i21]], i8* %m_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %fpcast.constant.fp.1.0, i8* %[[matA]], i8* %[[i44]], i8* %B, i8* %ldb_p, i8* %fpcast.constant.fp.0.0, i8* %[[i21]], i8* %m_p, i64 1, i64 1)
 ; CHECK:   %[[i45:.+]] = bitcast i64* %byref.constant.one.i to i8*
 ; CHECK:   %[[i46:.+]] = bitcast i64* %byref.mat.size.i to i8*
 ; CHECK:   store i64 1, i64* %byref.constant.one.i
@@ -207,13 +207,13 @@ entry:
 ; CHECK-NEXT:   %[[i62:.+]] = load double, double* %[[i61]]
 ; CHECK-NEXT:   %[[i63:.+]] = fadd fast double %[[i62]], %res.i
 ; CHECK-NEXT:   store double %[[i63]], double* %[[i61]]
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %byref.transpose.transb, i8* %m_p, i8* %k_p, i8* %n_p, i8* %alpha, i8* %"C'", i8* %ldc_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %"A'", i8* %lda_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %byref.transpose.transb, i8* %m_p, i8* %k_p, i8* %n_p, i8* %alpha, i8* %"C'", i8* %ldc_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %"A'", i8* %lda_p, i64 1, i64 1)
 ; CHECK-NEXT:   %loaded.trans8 = load i8, i8* %transa
 ; CHECK-DAG:   %[[i64:.+]] = icmp eq i8 %loaded.trans8, 78
 ; CHECK-DAG:   %[[i65:.+]] = icmp eq i8 %loaded.trans8, 110
 ; CHECK-DAG:   %[[i66:.+]] = or i1 %[[i65]], %[[i64]]
 ; CHECK-NEXT:   %[[i67:.+]] = select i1 %[[i66]], i8* %m_p, i8* %k_p
-; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %transb, i8* %k_p, i8* %n_p, i8* %m_p, i8* %alpha, i8* %[[matA]], i8* %[[i67]], i8* %"C'", i8* %ldc_p, i8* %beta, i8* %"B'", i8* %ldb_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %transb, i8* %k_p, i8* %n_p, i8* %m_p, i8* %alpha, i8* %[[matA]], i8* %[[i67]], i8* %"C'", i8* %ldc_p, i8* %beta, i8* %"B'", i8* %ldb_p, i64 1, i64 1)
 ; CHECK:   %[[i68:.+]] = bitcast i64* %byref.constant.one.i15 to i8*
 ; CHECK:   %[[i69:.+]] = bitcast i64* %byref.mat.size.i18 to i8*
 ; CHECK:   store i64 1, i64* %byref.constant.one.i15

diff --git a/enzyme/test/Enzyme/ReverseMode/blas/gemm_f_c_lacpy_runtime_act.ll b/enzyme/test/Enzyme/ReverseMode/blas/gemm_f_c_lacpy_runtime_act.ll
@@ -1,7 +1,7 @@
 ;RUN: if [ %llvmver -lt 16 ]; then %opt < %s %loadEnzyme -enzyme -enzyme-lapack-copy=1 -enzyme-runtime-activity=1 -S | FileCheck %s; fi
 ;RUN: %opt < %s %newLoadEnzyme -passes="enzyme" -enzyme-lapack-copy=1 -enzyme-runtime-activity=1 -S | FileCheck %s
 
-declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly) 
+declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i64, i64) 
 
 define void @f(i8* noalias %C, i8* noalias %A, i8* noalias %B, i8* noalias %alpha, i8* noalias %beta) {
 entry:
@@ -27,7 +27,7 @@ entry:
   store i64 4, i64* %lda, align 16
   store i64 8, i64* %ldb, align 16
   store i64 4, i64* %ldc, align 16
-  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %C, i8* %ldc_p) 
+  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %C, i8* %ldc_p, i64 1, i64 1) 
   %ptr = bitcast i8* %A to double*
   store double 0.0000000e+00, double* %ptr, align 8
   ret void
@@ -122,7 +122,7 @@ entry:
 ; CHECK-NEXT:   %malloccall6 = tail call noalias nonnull i8* @malloc(i64 %mallocsize5)
 ; CHECK-NEXT:   %mat_AB = bitcast i8* %malloccall6 to double*
 ; CHECK-NEXT:   %[[i21:.+]] = bitcast double* %mat_AB to i8*
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %C, i8* %ldc_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %C, i8* %ldc_p, i64 1, i64 1)
 ; CHECK-NEXT:   %"ptr'ipc" = bitcast i8* %"A'" to double*
 ; CHECK-NEXT:   %ptr = bitcast i8* %A to double*
 ; CHECK-NEXT:   store double 0.000000e+00, double* %ptr, align 8, !alias.scope !0, !noalias !3
@@ -167,7 +167,7 @@ entry:
 ; CHECK-NEXT:   %[[i44:.+]] = select i1 %[[i43]], i8* %m_p, i8* %k_p
 ; CHECK-NEXT:   store double 0.000000e+00, double* %byref.constant.fp.0.0
 ; CHECK-NEXT:   %fpcast.constant.fp.0.0 = bitcast double* %byref.constant.fp.0.0 to i8*
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %fpcast.constant.fp.1.0, i8* %[[matA]], i8* %[[i44]], i8* %B, i8* %ldb_p, i8* %fpcast.constant.fp.0.0, i8* %[[i21]], i8* %m_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %fpcast.constant.fp.1.0, i8* %[[matA]], i8* %[[i44]], i8* %B, i8* %ldb_p, i8* %fpcast.constant.fp.0.0, i8* %[[i21]], i8* %m_p, i64 1, i64 1)
 ; CHECK:   %[[i45:.+]] = bitcast i64* %byref.constant.one.i to i8*
 ; CHECK:   %[[i46:.+]] = bitcast i64* %byref.mat.size.i to i8*
 ; CHECK:   store i64 1, i64* %byref.constant.one.i
@@ -221,7 +221,7 @@ entry:
 ; CHECK-NEXT:   br i1 %rt.inactive.A, label %invertentry.A.done, label %invertentry.A.active
 
 ; CHECK: invertentry.A.active:                             ; preds = %invertentry.alpha.done
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %byref.transpose.transb, i8* %m_p, i8* %k_p, i8* %n_p, i8* %alpha, i8* %"C'", i8* %ldc_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %"A'", i8* %lda_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %byref.transpose.transb, i8* %m_p, i8* %k_p, i8* %n_p, i8* %alpha, i8* %"C'", i8* %ldc_p, i8* %B, i8* %ldb_p, i8* %beta, i8* %"A'", i8* %lda_p, i64 1, i64 1)
 ; CHECK-NEXT:   br label %invertentry.A.done
 
 ; CHECK: invertentry.A.done:                               ; preds = %invertentry.A.active, %invertentry.alpha.done
@@ -233,7 +233,7 @@ entry:
 ; CHECK-DAG:   %[[i65:.+]] = icmp eq i8 %loaded.trans8, 110
 ; CHECK-DAG:   %[[i66:.+]] = or i1 %[[i65]], %[[i64]]
 ; CHECK-NEXT:   %[[i67:.+]] = select i1 %[[i66]], i8* %m_p, i8* %k_p
-; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %transb, i8* %k_p, i8* %n_p, i8* %m_p, i8* %alpha, i8* %[[matA]], i8* %[[i67]], i8* %"C'", i8* %ldc_p, i8* %beta, i8* %"B'", i8* %ldb_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %transb, i8* %k_p, i8* %n_p, i8* %m_p, i8* %alpha, i8* %[[matA]], i8* %[[i67]], i8* %"C'", i8* %ldc_p, i8* %beta, i8* %"B'", i8* %ldb_p, i64 1, i64 1)
 ; CHECK-NEXT:   br label %invertentry.B.done
 
 ; CHECK: invertentry.B.done:                               ; preds = %invertentry.B.active, %invertentry.A.done

diff --git a/enzyme/test/Enzyme/ReverseMode/blas/gemm_f_c_loop.ll b/enzyme/test/Enzyme/ReverseMode/blas/gemm_f_c_loop.ll
@@ -1,7 +1,7 @@
 ;RUN: if [ %llvmver -lt 16 ]; then %opt < %s %loadEnzyme -enzyme -S | FileCheck %s; fi
 ;RUN: %opt < %s %newLoadEnzyme -passes="enzyme" -S | FileCheck %s
 
-declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly) 
+declare void @dgemm_64_(i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i8*, i8* nocapture readonly, i8* nocapture readonly, i8*, i8* nocapture readonly, i64, i64) 
 
 declare i8* @AData(i64)
 declare i8* @Aldap(i64)
@@ -43,7 +43,7 @@ loop:
   store i64 4, i64* %ldc, align 16
   %A = call i8* @AData(i64 %i) "enzyme_inactive"
   %lda_p = call i8* @Aldap(i64 %i) "enzyme_inactive"
-  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p) 
+  call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p, i64 1, i64 1) 
   call void @free(i8* %m_p)
   %cmp = icmp eq i64 %inc, 10
   br i1 %cmp, label %exit, label %loop
@@ -187,7 +187,7 @@ entry:
 ; CHECK-NEXT:   br i1 %23, label %__enzyme_memcpy_double_mat_64.exit, label %init.idx.i
 
 ; CHECK: __enzyme_memcpy_double_mat_64.exit:               ; preds = %loop, %init.end.i
-; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %transa, i8* %transb, i8* %m_p, i8* %n_p, i8* %k_p, i8* %alpha_p, i8* %A, i8* %lda_p, i8* %B, i8* %ldb_p, i8* %beta_p, i8* %C, i8* %ldc_p, i64 1, i64 1)
 ; CHECK-NEXT:   call void @free(i8* %m_p)
 ; CHECK-NEXT:   %cmp = icmp eq i64 %iv.next, 10
 ; CHECK-NEXT:   br i1 %cmp, label %exit, label %loop
@@ -275,7 +275,7 @@ entry:
 ; CHECK-DAG:   %[[r19:.+]] = icmp eq i8 %loaded.trans30, 110
 ; CHECK-DAG:   %[[r20:.+]] = or i1 %[[r19]], %[[r18]]
 ; CHECK-DAG:   %[[r21:.+]] = select i1 %[[r20]], i8* %[[i46]], i8* %cast.k
-; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %byref.transb, i8* %cast.k, i8* %n_p_unwrap, i8* %[[i46]], i8* %cast.alpha, i8* %[[i44]], i8* %[[r21]], i8* %"C'", i8* %cast.ldc, i8* %cast.beta, i8* %"B'", i8* %cast.ldb)
+; CHECK-NEXT:   call void @dgemm_64_(i8* %byref.transpose.transa, i8* %byref.transb, i8* %cast.k, i8* %n_p_unwrap, i8* %[[i46]], i8* %cast.alpha, i8* %[[i44]], i8* %[[r21]], i8* %"C'", i8* %cast.ldc, i8* %cast.beta, i8* %"B'", i8* %cast.ldb, i64 1, i64 1)
 ; CHECK-NEXT:   store i8 71, i8* %byref.constant.char.G
 ; CHECK-NEXT:   store i64 0, i64* %byref.constant.int.0
 ; CHECK-NEXT:   %intcast.constant.int.0 = bitcast i64* %byref.constant.int.0 to i8*