[llvm] [X86][GlobalISel] Support G_FADD, G_FSUB, G_FMUL, G_FDIV (PR #87339)
Evgenii Kudriashov via llvm-commits
llvm-commits at lists.llvm.org
Tue Apr 2 06:18:09 PDT 2024
https://github.com/e-kud updated https://github.com/llvm/llvm-project/pull/87339
>From 77c94380dc087ba83f91ab8950e5f4d7cde8c4d4 Mon Sep 17 00:00:00 2001
From: Evgenii Kudriashov <evgenii.kudriashov at intel.com>
Date: Tue, 2 Apr 2024 02:57:07 -0700
Subject: [PATCH 1/2] [X86][GlobalISel] Support G_FADD, G_FSUB, G_FMUL, G_FDIV
* Add vector legalization
* Add vector tests and reorganize existing tests
---
.../lib/Target/X86/GISel/X86LegalizerInfo.cpp | 7 +-
.../CodeGen/X86/GlobalISel/fadd-scalar.ll | 21 ---
.../CodeGen/X86/GlobalISel/fdiv-scalar.ll | 21 ---
.../CodeGen/X86/GlobalISel/fmul-scalar.ll | 21 ---
.../CodeGen/X86/GlobalISel/fsub-scalar.ll | 21 ---
llvm/test/CodeGen/X86/isel-fadd-vector.ll | 134 ++++++++++++++++++
llvm/test/CodeGen/X86/isel-fadd.ll | 52 +++++++
llvm/test/CodeGen/X86/isel-fdiv-vector.ll | 134 ++++++++++++++++++
llvm/test/CodeGen/X86/isel-fdiv.ll | 52 +++++++
llvm/test/CodeGen/X86/isel-fmul-vector.ll | 134 ++++++++++++++++++
llvm/test/CodeGen/X86/isel-fmul.ll | 52 +++++++
llvm/test/CodeGen/X86/isel-fsub-vector.ll | 134 ++++++++++++++++++
llvm/test/CodeGen/X86/isel-fsub.ll | 52 +++++++
13 files changed, 750 insertions(+), 85 deletions(-)
delete mode 100644 llvm/test/CodeGen/X86/GlobalISel/fadd-scalar.ll
delete mode 100644 llvm/test/CodeGen/X86/GlobalISel/fdiv-scalar.ll
delete mode 100644 llvm/test/CodeGen/X86/GlobalISel/fmul-scalar.ll
delete mode 100644 llvm/test/CodeGen/X86/GlobalISel/fsub-scalar.ll
create mode 100644 llvm/test/CodeGen/X86/isel-fadd-vector.ll
create mode 100644 llvm/test/CodeGen/X86/isel-fadd.ll
create mode 100644 llvm/test/CodeGen/X86/isel-fdiv-vector.ll
create mode 100644 llvm/test/CodeGen/X86/isel-fdiv.ll
create mode 100644 llvm/test/CodeGen/X86/isel-fmul-vector.ll
create mode 100644 llvm/test/CodeGen/X86/isel-fmul.ll
create mode 100644 llvm/test/CodeGen/X86/isel-fsub-vector.ll
create mode 100644 llvm/test/CodeGen/X86/isel-fsub.ll
diff --git a/llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp b/llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp
index 07041cc5b0491c..3be3121bab66f8 100644
--- a/llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp
+++ b/llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp
@@ -429,7 +429,12 @@ X86LegalizerInfo::X86LegalizerInfo(const X86Subtarget &STI,
(HasAVX && typeInSet(0, {v8s32, v4s64})(Query)) ||
(HasAVX512 && typeInSet(0, {v16s32, v8s64})(Query)) ||
(UseX87 && typeInSet(0, {s80})(Query));
- });
+ })
+ .clampMinNumElements(0, s32, 4)
+ .clampMinNumElements(0, s64, 2)
+ .clampMaxNumElements(0, s32, HasAVX512 ? 16 : (HasAVX ? 8 : 4))
+ .clampMaxNumElements(0, s64, HasAVX512 ? 8 : (HasAVX ? 4 : 2))
+ .scalarize(0);
// fp comparison
getActionDefinitionsBuilder(G_FCMP)
diff --git a/llvm/test/CodeGen/X86/GlobalISel/fadd-scalar.ll b/llvm/test/CodeGen/X86/GlobalISel/fadd-scalar.ll
deleted file mode 100644
index 73be29e5d8438b..00000000000000
--- a/llvm/test/CodeGen/X86/GlobalISel/fadd-scalar.ll
+++ /dev/null
@@ -1,21 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=x86_64-linux-gnu -global-isel -verify-machineinstrs < %s -o - | FileCheck %s --check-prefix=X64
-
-define float @test_fadd_float(float %arg1, float %arg2) {
-; X64-LABEL: test_fadd_float:
-; X64: # %bb.0:
-; X64-NEXT: addss %xmm1, %xmm0
-; X64-NEXT: retq
- %ret = fadd float %arg1, %arg2
- ret float %ret
-}
-
-define double @test_fadd_double(double %arg1, double %arg2) {
-; X64-LABEL: test_fadd_double:
-; X64: # %bb.0:
-; X64-NEXT: addsd %xmm1, %xmm0
-; X64-NEXT: retq
- %ret = fadd double %arg1, %arg2
- ret double %ret
-}
-
diff --git a/llvm/test/CodeGen/X86/GlobalISel/fdiv-scalar.ll b/llvm/test/CodeGen/X86/GlobalISel/fdiv-scalar.ll
deleted file mode 100644
index f2dc6de08528ed..00000000000000
--- a/llvm/test/CodeGen/X86/GlobalISel/fdiv-scalar.ll
+++ /dev/null
@@ -1,21 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=x86_64-linux-gnu -global-isel -verify-machineinstrs < %s -o - | FileCheck %s --check-prefix=X64
-
-define float @test_fdiv_float(float %arg1, float %arg2) {
-; X64-LABEL: test_fdiv_float:
-; X64: # %bb.0:
-; X64-NEXT: divss %xmm1, %xmm0
-; X64-NEXT: retq
- %ret = fdiv float %arg1, %arg2
- ret float %ret
-}
-
-define double @test_fdiv_double(double %arg1, double %arg2) {
-; X64-LABEL: test_fdiv_double:
-; X64: # %bb.0:
-; X64-NEXT: divsd %xmm1, %xmm0
-; X64-NEXT: retq
- %ret = fdiv double %arg1, %arg2
- ret double %ret
-}
-
diff --git a/llvm/test/CodeGen/X86/GlobalISel/fmul-scalar.ll b/llvm/test/CodeGen/X86/GlobalISel/fmul-scalar.ll
deleted file mode 100644
index 187593d082c543..00000000000000
--- a/llvm/test/CodeGen/X86/GlobalISel/fmul-scalar.ll
+++ /dev/null
@@ -1,21 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=x86_64-linux-gnu -global-isel -verify-machineinstrs < %s -o - | FileCheck %s --check-prefix=X64
-
-define float @test_fmul_float(float %arg1, float %arg2) {
-; X64-LABEL: test_fmul_float:
-; X64: # %bb.0:
-; X64-NEXT: mulss %xmm1, %xmm0
-; X64-NEXT: retq
- %ret = fmul float %arg1, %arg2
- ret float %ret
-}
-
-define double @test_fmul_double(double %arg1, double %arg2) {
-; X64-LABEL: test_fmul_double:
-; X64: # %bb.0:
-; X64-NEXT: mulsd %xmm1, %xmm0
-; X64-NEXT: retq
- %ret = fmul double %arg1, %arg2
- ret double %ret
-}
-
diff --git a/llvm/test/CodeGen/X86/GlobalISel/fsub-scalar.ll b/llvm/test/CodeGen/X86/GlobalISel/fsub-scalar.ll
deleted file mode 100644
index b2a82834af5bbe..00000000000000
--- a/llvm/test/CodeGen/X86/GlobalISel/fsub-scalar.ll
+++ /dev/null
@@ -1,21 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=x86_64-linux-gnu -global-isel -verify-machineinstrs < %s -o - | FileCheck %s --check-prefix=X64
-
-define float @test_fsub_float(float %arg1, float %arg2) {
-; X64-LABEL: test_fsub_float:
-; X64: # %bb.0:
-; X64-NEXT: subss %xmm1, %xmm0
-; X64-NEXT: retq
- %ret = fsub float %arg1, %arg2
- ret float %ret
-}
-
-define double @test_fsub_double(double %arg1, double %arg2) {
-; X64-LABEL: test_fsub_double:
-; X64: # %bb.0:
-; X64-NEXT: subsd %xmm1, %xmm0
-; X64-NEXT: retq
- %ret = fsub double %arg1, %arg2
- ret double %ret
-}
-
diff --git a/llvm/test/CodeGen/X86/isel-fadd-vector.ll b/llvm/test/CodeGen/X86/isel-fadd-vector.ll
new file mode 100644
index 00000000000000..b44818d49344db
--- /dev/null
+++ b/llvm/test/CodeGen/X86/isel-fadd-vector.ll
@@ -0,0 +1,134 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=AVX512
+
+define <4 x float> @test_fadd_v4s32(<4 x float> %arg1, <4 x float> %arg2) {
+; SSE-LABEL: test_fadd_v4s32:
+; SSE: # %bb.0:
+; SSE-NEXT: addps %xmm1, %xmm0
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fadd_v4s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vaddps %xmm1, %xmm0, %xmm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fadd_v4s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vaddps %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: retq
+ %ret = fadd <4 x float> %arg1, %arg2
+ ret <4 x float> %ret
+}
+
+define <2 x double> @test_fadd_v2s64(<2 x double> %arg1, <2 x double> %arg2) {
+; SSE-LABEL: test_fadd_v2s64:
+; SSE: # %bb.0:
+; SSE-NEXT: addpd %xmm1, %xmm0
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fadd_v2s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vaddpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fadd_v2s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vaddpd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: retq
+ %ret = fadd <2 x double> %arg1, %arg2
+ ret <2 x double> %ret
+}
+
+define <8 x float> @test_fadd_v8s32(<8 x float> %arg1, <8 x float> %arg2) {
+; SSE-LABEL: test_fadd_v8s32:
+; SSE: # %bb.0:
+; SSE-NEXT: addps %xmm2, %xmm0
+; SSE-NEXT: addps %xmm3, %xmm1
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fadd_v8s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vaddps %ymm1, %ymm0, %ymm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fadd_v8s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vaddps %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: retq
+ %ret = fadd <8 x float> %arg1, %arg2
+ ret <8 x float> %ret
+}
+
+define <4 x double> @test_fadd_v4s64(<4 x double> %arg1, <4 x double> %arg2) {
+; SSE-LABEL: test_fadd_v4s64:
+; SSE: # %bb.0:
+; SSE-NEXT: addpd %xmm2, %xmm0
+; SSE-NEXT: addpd %xmm3, %xmm1
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fadd_v4s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vaddpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fadd_v4s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vaddpd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: retq
+ %ret = fadd <4 x double> %arg1, %arg2
+ ret <4 x double> %ret
+}
+
+define <16 x float> @test_fadd_v16s32(<16 x float> %arg1, <16 x float> %arg2) {
+; SSE-LABEL: test_fadd_v16s32:
+; SSE: # %bb.0:
+; SSE-NEXT: addps %xmm4, %xmm0
+; SSE-NEXT: addps %xmm5, %xmm1
+; SSE-NEXT: addps %xmm6, %xmm2
+; SSE-NEXT: addps %xmm7, %xmm3
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fadd_v16s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vaddps %ymm2, %ymm0, %ymm0
+; AVX-NEXT: vaddps %ymm3, %ymm1, %ymm1
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fadd_v16s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vaddps %zmm1, %zmm0, %zmm0
+; AVX512-NEXT: retq
+ %ret = fadd <16 x float> %arg1, %arg2
+ ret <16 x float> %ret
+}
+
+define <8 x double> @test_fadd_v8s64(<8 x double> %arg1, <8 x double> %arg2) {
+; SSE-LABEL: test_fadd_v8s64:
+; SSE: # %bb.0:
+; SSE-NEXT: addpd %xmm4, %xmm0
+; SSE-NEXT: addpd %xmm5, %xmm1
+; SSE-NEXT: addpd %xmm6, %xmm2
+; SSE-NEXT: addpd %xmm7, %xmm3
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fadd_v8s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vaddpd %ymm2, %ymm0, %ymm0
+; AVX-NEXT: vaddpd %ymm3, %ymm1, %ymm1
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fadd_v8s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vaddpd %zmm1, %zmm0, %zmm0
+; AVX512-NEXT: retq
+ %ret = fadd <8 x double> %arg1, %arg2
+ ret <8 x double> %ret
+}
diff --git a/llvm/test/CodeGen/X86/isel-fadd.ll b/llvm/test/CodeGen/X86/isel-fadd.ll
new file mode 100644
index 00000000000000..16e005828608c1
--- /dev/null
+++ b/llvm/test/CodeGen/X86/isel-fadd.ll
@@ -0,0 +1,52 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-- | FileCheck %s --check-prefixes=SDAG-X86
+; RUN: llc < %s -mtriple=i686-- -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=FASTISEL-X86
+; TODO: Enable when x87 is supported
+; llc < %s -mtriple=i686-- -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=GISEL-X86
+; RUN: llc < %s -mtriple=x86_64-- | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=x86_64-- -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=x86_64-- -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=X64
+
+define float @test_fadd_float(float %arg1, float %arg2) {
+; SDAG-X86-LABEL: test_fadd_float:
+; SDAG-X86: # %bb.0:
+; SDAG-X86-NEXT: flds {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: fadds {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: retl
+;
+; FASTISEL-X86-LABEL: test_fadd_float:
+; FASTISEL-X86: # %bb.0:
+; FASTISEL-X86-NEXT: flds {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: flds {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: faddp %st, %st(1)
+; FASTISEL-X86-NEXT: retl
+;
+; X64-LABEL: test_fadd_float:
+; X64: # %bb.0:
+; X64-NEXT: addss %xmm1, %xmm0
+; X64-NEXT: retq
+ %ret = fadd float %arg1, %arg2
+ ret float %ret
+}
+
+define double @test_fadd_double(double %arg1, double %arg2) {
+; SDAG-X86-LABEL: test_fadd_double:
+; SDAG-X86: # %bb.0:
+; SDAG-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: faddl {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: retl
+;
+; FASTISEL-X86-LABEL: test_fadd_double:
+; FASTISEL-X86: # %bb.0:
+; FASTISEL-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: faddp %st, %st(1)
+; FASTISEL-X86-NEXT: retl
+;
+; X64-LABEL: test_fadd_double:
+; X64: # %bb.0:
+; X64-NEXT: addsd %xmm1, %xmm0
+; X64-NEXT: retq
+ %ret = fadd double %arg1, %arg2
+ ret double %ret
+}
diff --git a/llvm/test/CodeGen/X86/isel-fdiv-vector.ll b/llvm/test/CodeGen/X86/isel-fdiv-vector.ll
new file mode 100644
index 00000000000000..ff114c2af7e96b
--- /dev/null
+++ b/llvm/test/CodeGen/X86/isel-fdiv-vector.ll
@@ -0,0 +1,134 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=AVX512
+
+define <4 x float> @test_fdiv_v4s32(<4 x float> %arg1, <4 x float> %arg2) {
+; SSE-LABEL: test_fdiv_v4s32:
+; SSE: # %bb.0:
+; SSE-NEXT: divps %xmm1, %xmm0
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fdiv_v4s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vdivps %xmm1, %xmm0, %xmm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fdiv_v4s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vdivps %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: retq
+ %ret = fdiv <4 x float> %arg1, %arg2
+ ret <4 x float> %ret
+}
+
+define <2 x double> @test_fdiv_v2s64(<2 x double> %arg1, <2 x double> %arg2) {
+; SSE-LABEL: test_fdiv_v2s64:
+; SSE: # %bb.0:
+; SSE-NEXT: divpd %xmm1, %xmm0
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fdiv_v2s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vdivpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fdiv_v2s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vdivpd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: retq
+ %ret = fdiv <2 x double> %arg1, %arg2
+ ret <2 x double> %ret
+}
+
+define <8 x float> @test_fdiv_v8s32(<8 x float> %arg1, <8 x float> %arg2) {
+; SSE-LABEL: test_fdiv_v8s32:
+; SSE: # %bb.0:
+; SSE-NEXT: divps %xmm2, %xmm0
+; SSE-NEXT: divps %xmm3, %xmm1
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fdiv_v8s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vdivps %ymm1, %ymm0, %ymm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fdiv_v8s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vdivps %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: retq
+ %ret = fdiv <8 x float> %arg1, %arg2
+ ret <8 x float> %ret
+}
+
+define <4 x double> @test_fdiv_v4s64(<4 x double> %arg1, <4 x double> %arg2) {
+; SSE-LABEL: test_fdiv_v4s64:
+; SSE: # %bb.0:
+; SSE-NEXT: divpd %xmm2, %xmm0
+; SSE-NEXT: divpd %xmm3, %xmm1
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fdiv_v4s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vdivpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fdiv_v4s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vdivpd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: retq
+ %ret = fdiv <4 x double> %arg1, %arg2
+ ret <4 x double> %ret
+}
+
+define <16 x float> @test_fdiv_v16s32(<16 x float> %arg1, <16 x float> %arg2) {
+; SSE-LABEL: test_fdiv_v16s32:
+; SSE: # %bb.0:
+; SSE-NEXT: divps %xmm4, %xmm0
+; SSE-NEXT: divps %xmm5, %xmm1
+; SSE-NEXT: divps %xmm6, %xmm2
+; SSE-NEXT: divps %xmm7, %xmm3
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fdiv_v16s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vdivps %ymm2, %ymm0, %ymm0
+; AVX-NEXT: vdivps %ymm3, %ymm1, %ymm1
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fdiv_v16s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vdivps %zmm1, %zmm0, %zmm0
+; AVX512-NEXT: retq
+ %ret = fdiv <16 x float> %arg1, %arg2
+ ret <16 x float> %ret
+}
+
+define <8 x double> @test_fdiv_v8s64(<8 x double> %arg1, <8 x double> %arg2) {
+; SSE-LABEL: test_fdiv_v8s64:
+; SSE: # %bb.0:
+; SSE-NEXT: divpd %xmm4, %xmm0
+; SSE-NEXT: divpd %xmm5, %xmm1
+; SSE-NEXT: divpd %xmm6, %xmm2
+; SSE-NEXT: divpd %xmm7, %xmm3
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fdiv_v8s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vdivpd %ymm2, %ymm0, %ymm0
+; AVX-NEXT: vdivpd %ymm3, %ymm1, %ymm1
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fdiv_v8s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vdivpd %zmm1, %zmm0, %zmm0
+; AVX512-NEXT: retq
+ %ret = fdiv <8 x double> %arg1, %arg2
+ ret <8 x double> %ret
+}
diff --git a/llvm/test/CodeGen/X86/isel-fdiv.ll b/llvm/test/CodeGen/X86/isel-fdiv.ll
new file mode 100644
index 00000000000000..74ad27a8414d2c
--- /dev/null
+++ b/llvm/test/CodeGen/X86/isel-fdiv.ll
@@ -0,0 +1,52 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-- | FileCheck %s --check-prefixes=SDAG-X86
+; RUN: llc < %s -mtriple=i686-- -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=FASTISEL-X86
+; TODO: Enable when x87 is supported
+; llc < %s -mtriple=i686-- -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=GISEL-X86
+; RUN: llc < %s -mtriple=x86_64-- | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=x86_64-- -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=x86_64-- -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=X64
+
+define float @test_fdiv_float(float %arg1, float %arg2) {
+; SDAG-X86-LABEL: test_fdiv_float:
+; SDAG-X86: # %bb.0:
+; SDAG-X86-NEXT: flds {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: fdivs {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: retl
+;
+; FASTISEL-X86-LABEL: test_fdiv_float:
+; FASTISEL-X86: # %bb.0:
+; FASTISEL-X86-NEXT: flds {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: flds {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fdivp %st, %st(1)
+; FASTISEL-X86-NEXT: retl
+;
+; X64-LABEL: test_fdiv_float:
+; X64: # %bb.0:
+; X64-NEXT: divss %xmm1, %xmm0
+; X64-NEXT: retq
+ %ret = fdiv float %arg1, %arg2
+ ret float %ret
+}
+
+define double @test_fdiv_double(double %arg1, double %arg2) {
+; SDAG-X86-LABEL: test_fdiv_double:
+; SDAG-X86: # %bb.0:
+; SDAG-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: fdivl {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: retl
+;
+; FASTISEL-X86-LABEL: test_fdiv_double:
+; FASTISEL-X86: # %bb.0:
+; FASTISEL-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fdivp %st, %st(1)
+; FASTISEL-X86-NEXT: retl
+;
+; X64-LABEL: test_fdiv_double:
+; X64: # %bb.0:
+; X64-NEXT: divsd %xmm1, %xmm0
+; X64-NEXT: retq
+ %ret = fdiv double %arg1, %arg2
+ ret double %ret
+}
diff --git a/llvm/test/CodeGen/X86/isel-fmul-vector.ll b/llvm/test/CodeGen/X86/isel-fmul-vector.ll
new file mode 100644
index 00000000000000..7734f00f351cfb
--- /dev/null
+++ b/llvm/test/CodeGen/X86/isel-fmul-vector.ll
@@ -0,0 +1,134 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=AVX512
+
+define <4 x float> @test_fmul_v4s32(<4 x float> %arg1, <4 x float> %arg2) {
+; SSE-LABEL: test_fmul_v4s32:
+; SSE: # %bb.0:
+; SSE-NEXT: mulps %xmm1, %xmm0
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fmul_v4s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vmulps %xmm1, %xmm0, %xmm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fmul_v4s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vmulps %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: retq
+ %ret = fmul <4 x float> %arg1, %arg2
+ ret <4 x float> %ret
+}
+
+define <2 x double> @test_fmul_v2s64(<2 x double> %arg1, <2 x double> %arg2) {
+; SSE-LABEL: test_fmul_v2s64:
+; SSE: # %bb.0:
+; SSE-NEXT: mulpd %xmm1, %xmm0
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fmul_v2s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vmulpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fmul_v2s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vmulpd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: retq
+ %ret = fmul <2 x double> %arg1, %arg2
+ ret <2 x double> %ret
+}
+
+define <8 x float> @test_fmul_v8s32(<8 x float> %arg1, <8 x float> %arg2) {
+; SSE-LABEL: test_fmul_v8s32:
+; SSE: # %bb.0:
+; SSE-NEXT: mulps %xmm2, %xmm0
+; SSE-NEXT: mulps %xmm3, %xmm1
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fmul_v8s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vmulps %ymm1, %ymm0, %ymm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fmul_v8s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vmulps %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: retq
+ %ret = fmul <8 x float> %arg1, %arg2
+ ret <8 x float> %ret
+}
+
+define <4 x double> @test_fmul_v4s64(<4 x double> %arg1, <4 x double> %arg2) {
+; SSE-LABEL: test_fmul_v4s64:
+; SSE: # %bb.0:
+; SSE-NEXT: mulpd %xmm2, %xmm0
+; SSE-NEXT: mulpd %xmm3, %xmm1
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fmul_v4s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vmulpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fmul_v4s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vmulpd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: retq
+ %ret = fmul <4 x double> %arg1, %arg2
+ ret <4 x double> %ret
+}
+
+define <16 x float> @test_fmul_v16s32(<16 x float> %arg1, <16 x float> %arg2) {
+; SSE-LABEL: test_fmul_v16s32:
+; SSE: # %bb.0:
+; SSE-NEXT: mulps %xmm4, %xmm0
+; SSE-NEXT: mulps %xmm5, %xmm1
+; SSE-NEXT: mulps %xmm6, %xmm2
+; SSE-NEXT: mulps %xmm7, %xmm3
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fmul_v16s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vmulps %ymm2, %ymm0, %ymm0
+; AVX-NEXT: vmulps %ymm3, %ymm1, %ymm1
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fmul_v16s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vmulps %zmm1, %zmm0, %zmm0
+; AVX512-NEXT: retq
+ %ret = fmul <16 x float> %arg1, %arg2
+ ret <16 x float> %ret
+}
+
+define <8 x double> @test_fmul_v8s64(<8 x double> %arg1, <8 x double> %arg2) {
+; SSE-LABEL: test_fmul_v8s64:
+; SSE: # %bb.0:
+; SSE-NEXT: mulpd %xmm4, %xmm0
+; SSE-NEXT: mulpd %xmm5, %xmm1
+; SSE-NEXT: mulpd %xmm6, %xmm2
+; SSE-NEXT: mulpd %xmm7, %xmm3
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fmul_v8s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vmulpd %ymm2, %ymm0, %ymm0
+; AVX-NEXT: vmulpd %ymm3, %ymm1, %ymm1
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fmul_v8s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vmulpd %zmm1, %zmm0, %zmm0
+; AVX512-NEXT: retq
+ %ret = fmul <8 x double> %arg1, %arg2
+ ret <8 x double> %ret
+}
diff --git a/llvm/test/CodeGen/X86/isel-fmul.ll b/llvm/test/CodeGen/X86/isel-fmul.ll
new file mode 100644
index 00000000000000..af509a2967c2be
--- /dev/null
+++ b/llvm/test/CodeGen/X86/isel-fmul.ll
@@ -0,0 +1,52 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-- | FileCheck %s --check-prefixes=SDAG-X86
+; RUN: llc < %s -mtriple=i686-- -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=FASTISEL-X86
+; TODO: Enable when x87 is supported
+; llc < %s -mtriple=i686-- -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=GISEL-X86
+; RUN: llc < %s -mtriple=x86_64-- | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=x86_64-- -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=x86_64-- -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=X64
+
+define float @test_fmul_float(float %arg1, float %arg2) {
+; SDAG-X86-LABEL: test_fmul_float:
+; SDAG-X86: # %bb.0:
+; SDAG-X86-NEXT: flds {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: fmuls {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: retl
+;
+; FASTISEL-X86-LABEL: test_fmul_float:
+; FASTISEL-X86: # %bb.0:
+; FASTISEL-X86-NEXT: flds {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: flds {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fmulp %st, %st(1)
+; FASTISEL-X86-NEXT: retl
+;
+; X64-LABEL: test_fmul_float:
+; X64: # %bb.0:
+; X64-NEXT: mulss %xmm1, %xmm0
+; X64-NEXT: retq
+ %ret = fmul float %arg1, %arg2
+ ret float %ret
+}
+
+define double @test_fmul_double(double %arg1, double %arg2) {
+; SDAG-X86-LABEL: test_fmul_double:
+; SDAG-X86: # %bb.0:
+; SDAG-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: fmull {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: retl
+;
+; FASTISEL-X86-LABEL: test_fmul_double:
+; FASTISEL-X86: # %bb.0:
+; FASTISEL-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fmulp %st, %st(1)
+; FASTISEL-X86-NEXT: retl
+;
+; X64-LABEL: test_fmul_double:
+; X64: # %bb.0:
+; X64-NEXT: mulsd %xmm1, %xmm0
+; X64-NEXT: retq
+ %ret = fmul double %arg1, %arg2
+ ret double %ret
+}
diff --git a/llvm/test/CodeGen/X86/isel-fsub-vector.ll b/llvm/test/CodeGen/X86/isel-fsub-vector.ll
new file mode 100644
index 00000000000000..9e7a14bcdae561
--- /dev/null
+++ b/llvm/test/CodeGen/X86/isel-fsub-vector.ll
@@ -0,0 +1,134 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=SSE
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=AVX
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=AVX512
+
+define <4 x float> @test_fsub_v4s32(<4 x float> %arg1, <4 x float> %arg2) {
+; SSE-LABEL: test_fsub_v4s32:
+; SSE: # %bb.0:
+; SSE-NEXT: subps %xmm1, %xmm0
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fsub_v4s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vsubps %xmm1, %xmm0, %xmm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fsub_v4s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vsubps %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: retq
+ %ret = fsub <4 x float> %arg1, %arg2
+ ret <4 x float> %ret
+}
+
+define <2 x double> @test_fsub_v2s64(<2 x double> %arg1, <2 x double> %arg2) {
+; SSE-LABEL: test_fsub_v2s64:
+; SSE: # %bb.0:
+; SSE-NEXT: subpd %xmm1, %xmm0
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fsub_v2s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vsubpd %xmm1, %xmm0, %xmm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fsub_v2s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vsubpd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: retq
+ %ret = fsub <2 x double> %arg1, %arg2
+ ret <2 x double> %ret
+}
+
+define <8 x float> @test_fsub_v8s32(<8 x float> %arg1, <8 x float> %arg2) {
+; SSE-LABEL: test_fsub_v8s32:
+; SSE: # %bb.0:
+; SSE-NEXT: subps %xmm2, %xmm0
+; SSE-NEXT: subps %xmm3, %xmm1
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fsub_v8s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vsubps %ymm1, %ymm0, %ymm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fsub_v8s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vsubps %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: retq
+ %ret = fsub <8 x float> %arg1, %arg2
+ ret <8 x float> %ret
+}
+
+define <4 x double> @test_fsub_v4s64(<4 x double> %arg1, <4 x double> %arg2) {
+; SSE-LABEL: test_fsub_v4s64:
+; SSE: # %bb.0:
+; SSE-NEXT: subpd %xmm2, %xmm0
+; SSE-NEXT: subpd %xmm3, %xmm1
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fsub_v4s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vsubpd %ymm1, %ymm0, %ymm0
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fsub_v4s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vsubpd %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: retq
+ %ret = fsub <4 x double> %arg1, %arg2
+ ret <4 x double> %ret
+}
+
+define <16 x float> @test_fsub_v16s32(<16 x float> %arg1, <16 x float> %arg2) {
+; SSE-LABEL: test_fsub_v16s32:
+; SSE: # %bb.0:
+; SSE-NEXT: subps %xmm4, %xmm0
+; SSE-NEXT: subps %xmm5, %xmm1
+; SSE-NEXT: subps %xmm6, %xmm2
+; SSE-NEXT: subps %xmm7, %xmm3
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fsub_v16s32:
+; AVX: # %bb.0:
+; AVX-NEXT: vsubps %ymm2, %ymm0, %ymm0
+; AVX-NEXT: vsubps %ymm3, %ymm1, %ymm1
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fsub_v16s32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vsubps %zmm1, %zmm0, %zmm0
+; AVX512-NEXT: retq
+ %ret = fsub <16 x float> %arg1, %arg2
+ ret <16 x float> %ret
+}
+
+define <8 x double> @test_fsub_v8s64(<8 x double> %arg1, <8 x double> %arg2) {
+; SSE-LABEL: test_fsub_v8s64:
+; SSE: # %bb.0:
+; SSE-NEXT: subpd %xmm4, %xmm0
+; SSE-NEXT: subpd %xmm5, %xmm1
+; SSE-NEXT: subpd %xmm6, %xmm2
+; SSE-NEXT: subpd %xmm7, %xmm3
+; SSE-NEXT: retq
+;
+; AVX-LABEL: test_fsub_v8s64:
+; AVX: # %bb.0:
+; AVX-NEXT: vsubpd %ymm2, %ymm0, %ymm0
+; AVX-NEXT: vsubpd %ymm3, %ymm1, %ymm1
+; AVX-NEXT: retq
+;
+; AVX512-LABEL: test_fsub_v8s64:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vsubpd %zmm1, %zmm0, %zmm0
+; AVX512-NEXT: retq
+ %ret = fsub <8 x double> %arg1, %arg2
+ ret <8 x double> %ret
+}
diff --git a/llvm/test/CodeGen/X86/isel-fsub.ll b/llvm/test/CodeGen/X86/isel-fsub.ll
new file mode 100644
index 00000000000000..943e633b0d9eb4
--- /dev/null
+++ b/llvm/test/CodeGen/X86/isel-fsub.ll
@@ -0,0 +1,52 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-- | FileCheck %s --check-prefixes=SDAG-X86
+; RUN: llc < %s -mtriple=i686-- -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=FASTISEL-X86
+; TODO: Enable when x87 is supported
+; llc < %s -mtriple=i686-- -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=GISEL-X86
+; RUN: llc < %s -mtriple=x86_64-- | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=x86_64-- -fast-isel -fast-isel-abort=1 | FileCheck %s --check-prefixes=X64
+; RUN: llc < %s -mtriple=x86_64-- -global-isel -global-isel-abort=1 | FileCheck %s --check-prefixes=X64
+
+define float @test_fsub_float(float %arg1, float %arg2) {
+; SDAG-X86-LABEL: test_fsub_float:
+; SDAG-X86: # %bb.0:
+; SDAG-X86-NEXT: flds {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: fsubs {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: retl
+;
+; FASTISEL-X86-LABEL: test_fsub_float:
+; FASTISEL-X86: # %bb.0:
+; FASTISEL-X86-NEXT: flds {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: flds {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fsubp %st, %st(1)
+; FASTISEL-X86-NEXT: retl
+;
+; X64-LABEL: test_fsub_float:
+; X64: # %bb.0:
+; X64-NEXT: subss %xmm1, %xmm0
+; X64-NEXT: retq
+ %ret = fsub float %arg1, %arg2
+ ret float %ret
+}
+
+define double @test_fsub_double(double %arg1, double %arg2) {
+; SDAG-X86-LABEL: test_fsub_double:
+; SDAG-X86: # %bb.0:
+; SDAG-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: fsubl {{[0-9]+}}(%esp)
+; SDAG-X86-NEXT: retl
+;
+; FASTISEL-X86-LABEL: test_fsub_double:
+; FASTISEL-X86: # %bb.0:
+; FASTISEL-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fldl {{[0-9]+}}(%esp)
+; FASTISEL-X86-NEXT: fsubp %st, %st(1)
+; FASTISEL-X86-NEXT: retl
+;
+; X64-LABEL: test_fsub_double:
+; X64: # %bb.0:
+; X64-NEXT: subsd %xmm1, %xmm0
+; X64-NEXT: retq
+ %ret = fsub double %arg1, %arg2
+ ret double %ret
+}
>From c9c6660d9f83dcb4848052d236f85ea634be0a47 Mon Sep 17 00:00:00 2001
From: Evgenii Kudriashov <evgenii.kudriashov at intel.com>
Date: Tue, 2 Apr 2024 06:14:27 -0700
Subject: [PATCH 2/2] Use clampNumElements instead of clampMin/Max
---
llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp | 6 ++----
1 file changed, 2 insertions(+), 4 deletions(-)
diff --git a/llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp b/llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp
index 3be3121bab66f8..631c6903743b39 100644
--- a/llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp
+++ b/llvm/lib/Target/X86/GISel/X86LegalizerInfo.cpp
@@ -430,10 +430,8 @@ X86LegalizerInfo::X86LegalizerInfo(const X86Subtarget &STI,
(HasAVX512 && typeInSet(0, {v16s32, v8s64})(Query)) ||
(UseX87 && typeInSet(0, {s80})(Query));
})
- .clampMinNumElements(0, s32, 4)
- .clampMinNumElements(0, s64, 2)
- .clampMaxNumElements(0, s32, HasAVX512 ? 16 : (HasAVX ? 8 : 4))
- .clampMaxNumElements(0, s64, HasAVX512 ? 8 : (HasAVX ? 4 : 2))
+ .clampNumElements(0, v4s32, HasAVX512 ? v16s32 : (HasAVX ? v8s32 : v4s32))
+ .clampNumElements(0, v2s64, HasAVX512 ? v8s64 : (HasAVX ? v4s64 : v2s64))
.scalarize(0);
// fp comparison
More information about the llvm-commits
mailing list