diff options
| author | Dale Johannesen <dalej@apple.com> | 2010-09-30 23:57:10 +0000 |
|---|---|---|
| committer | Dale Johannesen <dalej@apple.com> | 2010-09-30 23:57:10 +0000 |
| commit | 0488fb649a56b7fc89a5814df5308813f9e5a85d (patch) | |
| tree | 21913bb81960866b73c6b49f29782c142563304c /test/CodeGen/X86 | |
| parent | a7e3b564773cc45a1b08383c09fe86b17b3ffe92 (diff) | |
| download | llvm-0488fb649a56b7fc89a5814df5308813f9e5a85d.tar.gz llvm-0488fb649a56b7fc89a5814df5308813f9e5a85d.tar.bz2 llvm-0488fb649a56b7fc89a5814df5308813f9e5a85d.tar.xz | |
Massive rewrite of MMX:
The x86_mmx type is used for MMX intrinsics, parameters and
return values where these use MMX registers, and is also
supported in load, store, and bitcast.
Only the above operations generate MMX instructions, and optimizations
do not operate on or produce MMX intrinsics.
MMX-sized vectors <2 x i32> etc. are lowered to XMM or split into
smaller pieces. Optimizations may occur on these forms and the
result casted back to x86_mmx, provided the result feeds into a
previous existing x86_mmx operation.
The point of all this is prevent optimizations from introducing
MMX operations, which is unsafe due to the EMMS problem.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115243 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'test/CodeGen/X86')
23 files changed, 539 insertions, 213 deletions
diff --git a/test/CodeGen/X86/2007-04-25-MMX-PADDQ.ll b/test/CodeGen/X86/2007-04-25-MMX-PADDQ.ll index c39b82a1fe..256cbbb5d9 100644 --- a/test/CodeGen/X86/2007-04-25-MMX-PADDQ.ll +++ b/test/CodeGen/X86/2007-04-25-MMX-PADDQ.ll @@ -1,4 +1,5 @@ ; RUN: llc < %s -o - -march=x86 -mattr=+mmx | FileCheck %s +; There are no MMX instructions here. We use add+adcl for the adds. define <1 x i64> @unsigned_add3(<1 x i64>* %a, <1 x i64>* %b, i32 %count) nounwind { entry: @@ -7,9 +8,8 @@ entry: bb26: ; preds = %bb26, %entry -; CHECK: movq ({{.*}},8), %mm -; CHECK: paddq ({{.*}},8), %mm -; CHECK: paddq %mm{{[0-7]}}, %mm +; CHECK: addl %eax, %ebx +; CHECK: adcl %edx, %ebp %i.037.0 = phi i32 [ 0, %entry ], [ %tmp25, %bb26 ] ; <i32> [#uses=3] %sum.035.0 = phi <1 x i64> [ zeroinitializer, %entry ], [ %tmp22, %bb26 ] ; <<1 x i64>> [#uses=1] @@ -27,3 +27,38 @@ bb31: ; preds = %bb26, %entry %sum.035.1 = phi <1 x i64> [ zeroinitializer, %entry ], [ %tmp22, %bb26 ] ; <<1 x i64>> [#uses=1] ret <1 x i64> %sum.035.1 } + + +; This is the original test converted to use MMX intrinsics. + +define <1 x i64> @unsigned_add3a(x86_mmx* %a, x86_mmx* %b, i32 %count) nounwind { +entry: + %tmp2943 = bitcast <1 x i64><i64 0> to x86_mmx + %tmp2942 = icmp eq i32 %count, 0 ; <i1> [#uses=1] + br i1 %tmp2942, label %bb31, label %bb26 + +bb26: ; preds = %bb26, %entry + +; CHECK: movq ({{.*}},8), %mm +; CHECK: paddq ({{.*}},8), %mm +; CHECK: paddq %mm{{[0-7]}}, %mm + + %i.037.0 = phi i32 [ 0, %entry ], [ %tmp25, %bb26 ] ; <i32> [#uses=3] + %sum.035.0 = phi x86_mmx [ %tmp2943, %entry ], [ %tmp22, %bb26 ] ; <x86_mmx> [#uses=1] + %tmp13 = getelementptr x86_mmx* %b, i32 %i.037.0 ; <x86_mmx*> [#uses=1] + %tmp14 = load x86_mmx* %tmp13 ; <x86_mmx> [#uses=1] + %tmp18 = getelementptr x86_mmx* %a, i32 %i.037.0 ; <x86_mmx*> [#uses=1] + %tmp19 = load x86_mmx* %tmp18 ; <x86_mmx> [#uses=1] + %tmp21 = call x86_mmx @llvm.x86.mmx.padd.q (x86_mmx %tmp19, x86_mmx %tmp14) ; <x86_mmx> [#uses=1] + %tmp22 = call x86_mmx @llvm.x86.mmx.padd.q (x86_mmx %tmp21, x86_mmx %sum.035.0) ; <x86_mmx> [#uses=2] + %tmp25 = add i32 %i.037.0, 1 ; <i32> [#uses=2] + %tmp29 = icmp ult i32 %tmp25, %count ; <i1> [#uses=1] + br i1 %tmp29, label %bb26, label %bb31 + +bb31: ; preds = %bb26, %entry + %sum.035.1 = phi x86_mmx [ %tmp2943, %entry ], [ %tmp22, %bb26 ] ; <x86_mmx> [#uses=1] + %t = bitcast x86_mmx %sum.035.1 to <1 x i64> + ret <1 x i64> %t +} + +declare x86_mmx @llvm.x86.mmx.padd.q(x86_mmx, x86_mmx) diff --git a/test/CodeGen/X86/2007-05-15-maskmovq.ll b/test/CodeGen/X86/2007-05-15-maskmovq.ll index 2093b8f687..006cf2e43a 100644 --- a/test/CodeGen/X86/2007-05-15-maskmovq.ll +++ b/test/CodeGen/X86/2007-05-15-maskmovq.ll @@ -5,10 +5,10 @@ target triple = "i686-apple-darwin8" define void @test(<1 x i64> %c64, <1 x i64> %mask1, i8* %P) { entry: - %tmp4 = bitcast <1 x i64> %mask1 to <8 x i8> ; <<8 x i8>> [#uses=1] - %tmp6 = bitcast <1 x i64> %c64 to <8 x i8> ; <<8 x i8>> [#uses=1] - tail call void @llvm.x86.mmx.maskmovq( <8 x i8> %tmp6, <8 x i8> %tmp4, i8* %P ) + %tmp4 = bitcast <1 x i64> %mask1 to x86_mmx ; <x86_mmx> [#uses=1] + %tmp6 = bitcast <1 x i64> %c64 to x86_mmx ; <x86_mmx> [#uses=1] + tail call void @llvm.x86.mmx.maskmovq( x86_mmx %tmp4, x86_mmx %tmp6, i8* %P ) ret void } -declare void @llvm.x86.mmx.maskmovq(<8 x i8>, <8 x i8>, i8*) +declare void @llvm.x86.mmx.maskmovq(x86_mmx, x86_mmx, i8*) diff --git a/test/CodeGen/X86/2007-06-15-IntToMMX.ll b/test/CodeGen/X86/2007-06-15-IntToMMX.ll index 6128d8b92d..660d4fe7b1 100644 --- a/test/CodeGen/X86/2007-06-15-IntToMMX.ll +++ b/test/CodeGen/X86/2007-06-15-IntToMMX.ll @@ -1,17 +1,16 @@ ; RUN: llc < %s -march=x86-64 -mattr=+mmx | grep paddusw -@R = external global <1 x i64> ; <<1 x i64>*> [#uses=1] +@R = external global x86_mmx ; <x86_mmx*> [#uses=1] define void @foo(<1 x i64> %A, <1 x i64> %B) { entry: - %tmp4 = bitcast <1 x i64> %B to <4 x i16> ; <<4 x i16>> [#uses=1] - %tmp6 = bitcast <1 x i64> %A to <4 x i16> ; <<4 x i16>> [#uses=1] - %tmp7 = tail call <4 x i16> @llvm.x86.mmx.paddus.w( <4 x i16> %tmp6, <4 x i16> %tmp4 ) ; <<4 x i16>> [#uses=1] - %tmp8 = bitcast <4 x i16> %tmp7 to <1 x i64> ; <<1 x i64>> [#uses=1] - store <1 x i64> %tmp8, <1 x i64>* @R + %tmp2 = bitcast <1 x i64> %A to x86_mmx + %tmp3 = bitcast <1 x i64> %B to x86_mmx + %tmp7 = tail call x86_mmx @llvm.x86.mmx.paddus.w( x86_mmx %tmp2, x86_mmx %tmp3 ) ; <x86_mmx> [#uses=1] + store x86_mmx %tmp7, x86_mmx* @R tail call void @llvm.x86.mmx.emms( ) ret void } -declare <4 x i16> @llvm.x86.mmx.paddus.w(<4 x i16>, <4 x i16>) +declare x86_mmx @llvm.x86.mmx.paddus.w(x86_mmx, x86_mmx) declare void @llvm.x86.mmx.emms() diff --git a/test/CodeGen/X86/2007-07-03-GR64ToVR64.ll b/test/CodeGen/X86/2007-07-03-GR64ToVR64.ll index 2c513f1781..1c5e6766fd 100644 --- a/test/CodeGen/X86/2007-07-03-GR64ToVR64.ll +++ b/test/CodeGen/X86/2007-07-03-GR64ToVR64.ll @@ -2,19 +2,17 @@ ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+mmx | grep {movd %rdi, %mm1} ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+mmx | grep {paddusw %mm0, %mm1} -@R = external global <1 x i64> ; <<1 x i64>*> [#uses=1] +@R = external global x86_mmx ; <x86_mmx*> [#uses=1] define void @foo(<1 x i64> %A, <1 x i64> %B) nounwind { entry: - %tmp4 = bitcast <1 x i64> %B to <4 x i16> ; <<4 x i16>> [#uses=1] - %tmp6 = bitcast <1 x i64> %A to <4 x i16> ; <<4 x i16>> [#uses=1] - %tmp7 = tail call <4 x i16> @llvm.x86.mmx.paddus.w( <4 x i16> %tmp6, <4 x i16> %tmp4 ) ; <<4 x i16>> [#uses=1] - %tmp8 = bitcast <4 x i16> %tmp7 to <1 x i64> ; <<1 x i64>> [#uses=1] - store <1 x i64> %tmp8, <1 x i64>* @R + %tmp4 = bitcast <1 x i64> %B to x86_mmx ; <<4 x i16>> [#uses=1] + %tmp6 = bitcast <1 x i64> %A to x86_mmx ; <<4 x i16>> [#uses=1] + %tmp7 = tail call x86_mmx @llvm.x86.mmx.paddus.w( x86_mmx %tmp6, x86_mmx %tmp4 ) ; <x86_mmx> [#uses=1] + store x86_mmx %tmp7, x86_mmx* @R tail call void @llvm.x86.mmx.emms( ) ret void } -declare <4 x i16> @llvm.x86.mmx.paddus.w(<4 x i16>, <4 x i16>) - +declare x86_mmx @llvm.x86.mmx.paddus.w(x86_mmx, x86_mmx) declare void @llvm.x86.mmx.emms() diff --git a/test/CodeGen/X86/2008-04-08-CoalescerCrash.ll b/test/CodeGen/X86/2008-04-08-CoalescerCrash.ll index dc8c097efc..5089e8c5b6 100644 --- a/test/CodeGen/X86/2008-04-08-CoalescerCrash.ll +++ b/test/CodeGen/X86/2008-04-08-CoalescerCrash.ll @@ -5,15 +5,15 @@ entry: tail call void asm sideeffect "# top of block", "~{dirflag},~{fpsr},~{flags},~{di},~{si},~{dx},~{cx},~{ax}"( ) nounwind tail call void asm sideeffect ".file \224443946.c\22", "~{dirflag},~{fpsr},~{flags}"( ) nounwind tail call void asm sideeffect ".line 8", "~{dirflag},~{fpsr},~{flags}"( ) nounwind - %tmp1 = tail call <2 x i32> asm sideeffect "movd $1, $0", "=={mm4},{bp},~{dirflag},~{fpsr},~{flags},~{memory}"( i32 undef ) nounwind ; <<2 x i32>> [#uses=1] + %tmp1 = tail call x86_mmx asm sideeffect "movd $1, $0", "=={mm4},{bp},~{dirflag},~{fpsr},~{flags},~{memory}"( i32 undef ) nounwind ; <x86_mmx> [#uses=1] tail call void asm sideeffect ".file \224443946.c\22", "~{dirflag},~{fpsr},~{flags}"( ) nounwind tail call void asm sideeffect ".line 9", "~{dirflag},~{fpsr},~{flags}"( ) nounwind - %tmp3 = tail call i32 asm sideeffect "movd $1, $0", "=={bp},{mm3},~{dirflag},~{fpsr},~{flags},~{memory}"( <2 x i32> undef ) nounwind ; <i32> [#uses=1] + %tmp3 = tail call i32 asm sideeffect "movd $1, $0", "=={bp},{mm3},~{dirflag},~{fpsr},~{flags},~{memory}"( x86_mmx undef ) nounwind ; <i32> [#uses=1] tail call void asm sideeffect ".file \224443946.c\22", "~{dirflag},~{fpsr},~{flags}"( ) nounwind tail call void asm sideeffect ".line 10", "~{dirflag},~{fpsr},~{flags}"( ) nounwind - tail call void asm sideeffect "movntq $0, 0($1,$2)", "{mm0},{di},{bp},~{dirflag},~{fpsr},~{flags},~{memory}"( <2 x i32> undef, i32 undef, i32 %tmp3 ) nounwind + tail call void asm sideeffect "movntq $0, 0($1,$2)", "{mm0},{di},{bp},~{dirflag},~{fpsr},~{flags},~{memory}"( x86_mmx undef, i32 undef, i32 %tmp3 ) nounwind tail call void asm sideeffect ".file \224443946.c\22", "~{dirflag},~{fpsr},~{flags}"( ) nounwind tail call void asm sideeffect ".line 11", "~{dirflag},~{fpsr},~{flags}"( ) nounwind - %tmp8 = tail call i32 asm sideeffect "movd $1, $0", "=={bp},{mm4},~{dirflag},~{fpsr},~{flags},~{memory}"( <2 x i32> %tmp1 ) nounwind ; <i32> [#uses=0] + %tmp8 = tail call i32 asm sideeffect "movd $1, $0", "=={bp},{mm4},~{dirflag},~{fpsr},~{flags},~{memory}"( x86_mmx %tmp1 ) nounwind ; <i32> [#uses=0] ret i32 undef } diff --git a/test/CodeGen/X86/2008-08-23-64Bit-maskmovq.ll b/test/CodeGen/X86/2008-08-23-64Bit-maskmovq.ll index c76dd7de12..53402c0451 100644 --- a/test/CodeGen/X86/2008-08-23-64Bit-maskmovq.ll +++ b/test/CodeGen/X86/2008-08-23-64Bit-maskmovq.ll @@ -17,11 +17,13 @@ entry: br i1 false, label %bb.nph144.split, label %bb133 bb.nph144.split: ; preds = %entry - tail call void @llvm.x86.mmx.maskmovq( <8 x i8> zeroinitializer, <8 x i8> zeroinitializer, i8* null ) nounwind + %tmp = bitcast <8 x i8> zeroinitializer to x86_mmx + %tmp2 = bitcast <8 x i8> zeroinitializer to x86_mmx + tail call void @llvm.x86.mmx.maskmovq( x86_mmx %tmp, x86_mmx %tmp2, i8* null ) nounwind unreachable bb133: ; preds = %entry ret void } -declare void @llvm.x86.mmx.maskmovq(<8 x i8>, <8 x i8>, i8*) nounwind +declare void @llvm.x86.mmx.maskmovq(x86_mmx, x86_mmx, i8*) nounwind diff --git a/test/CodeGen/X86/2008-09-05-sinttofp-2xi32.ll b/test/CodeGen/X86/2008-09-05-sinttofp-2xi32.ll index 60be0d51e7..2dc1deaf17 100644 --- a/test/CodeGen/X86/2008-09-05-sinttofp-2xi32.ll +++ b/test/CodeGen/X86/2008-09-05-sinttofp-2xi32.ll @@ -1,6 +1,9 @@ +; RUN: llc < %s -march=x86 -mattr=+sse2 -mattr=+mmx | grep unpcklpd +; RUN: llc < %s -march=x86 -mattr=+sse2 -mattr=+mmx | grep unpckhpd ; RUN: llc < %s -march=x86 -mattr=+sse2 | grep cvttpd2pi | count 1 ; RUN: llc < %s -march=x86 -mattr=+sse2 | grep cvtpi2pd | count 1 -; PR2687 +; originally from PR2687, but things don't work that way any more. +; there are no MMX instructions here; we use XMM. define <2 x double> @a(<2 x i32> %x) nounwind { entry: @@ -13,3 +16,20 @@ entry: %y = fptosi <2 x double> %x to <2 x i32> ret <2 x i32> %y } + +; This is how to get MMX instructions. + +define <2 x double> @a2(x86_mmx %x) nounwind { +entry: + %y = tail call <2 x double> @llvm.x86.sse.cvtpi2pd(x86_mmx %x) + ret <2 x double> %y +} + +define x86_mmx @b2(<2 x double> %x) nounwind { +entry: + %y = tail call x86_mmx @llvm.x86.sse.cvttpd2pi (<2 x double> %x) + ret x86_mmx %y +} + +declare <2 x double> @llvm.x86.sse.cvtpi2pd(x86_mmx) +declare x86_mmx @llvm.x86.sse.cvttpd2pi(<2 x double>) diff --git a/test/CodeGen/X86/2009-08-02-mmx-scalar-to-vector.ll b/test/CodeGen/X86/2009-08-02-mmx-scalar-to-vector.ll index b9b09a3f00..288eef4f69 100644 --- a/test/CodeGen/X86/2009-08-02-mmx-scalar-to-vector.ll +++ b/test/CodeGen/X86/2009-08-02-mmx-scalar-to-vector.ll @@ -1,10 +1,12 @@ ; RUN: llc < %s -march=x86-64 ; PR4669 -declare <1 x i64> @llvm.x86.mmx.pslli.q(<1 x i64>, i32) +declare x86_mmx @llvm.x86.mmx.pslli.q(x86_mmx, i32) define <1 x i64> @test(i64 %t) { entry: %t1 = insertelement <1 x i64> undef, i64 %t, i32 0 - %t2 = tail call <1 x i64> @llvm.x86.mmx.pslli.q(<1 x i64> %t1, i32 48) - ret <1 x i64> %t2 + %t0 = bitcast <1 x i64> %t1 to x86_mmx + %t2 = tail call x86_mmx @llvm.x86.mmx.pslli.q(x86_mmx %t0, i32 48) + %t3 = bitcast x86_mmx %t2 to <1 x i64> + ret <1 x i64> %t3 } diff --git a/test/CodeGen/X86/2010-04-23-mmx-movdq2q.ll b/test/CodeGen/X86/2010-04-23-mmx-movdq2q.ll index 4cd3be35e8..fa3d5fbcdc 100644 --- a/test/CodeGen/X86/2010-04-23-mmx-movdq2q.ll +++ b/test/CodeGen/X86/2010-04-23-mmx-movdq2q.ll @@ -1,12 +1,12 @@ ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | FileCheck %s +; There are no MMX operations here, so we use XMM or i64. define void @ti8(double %a, double %b) nounwind { entry: %tmp1 = bitcast double %a to <8 x i8> -; CHECK: movdq2q %tmp2 = bitcast double %b to <8 x i8> -; CHECK: movdq2q %tmp3 = add <8 x i8> %tmp1, %tmp2 +; CHECK: paddb %xmm1, %xmm0 store <8 x i8> %tmp3, <8 x i8>* null ret void } @@ -14,10 +14,9 @@ entry: define void @ti16(double %a, double %b) nounwind { entry: %tmp1 = bitcast double %a to <4 x i16> -; CHECK: movdq2q %tmp2 = bitcast double %b to <4 x i16> -; CHECK: movdq2q %tmp3 = add <4 x i16> %tmp1, %tmp2 +; CHECK: paddw %xmm1, %xmm0 store <4 x i16> %tmp3, <4 x i16>* null ret void } @@ -25,10 +24,9 @@ entry: define void @ti32(double %a, double %b) nounwind { entry: %tmp1 = bitcast double %a to <2 x i32> -; CHECK: movdq2q %tmp2 = bitcast double %b to <2 x i32> -; CHECK: movdq2q %tmp3 = add <2 x i32> %tmp1, %tmp2 +; CHECK: paddd %xmm1, %xmm0 store <2 x i32> %tmp3, <2 x i32>* null ret void } @@ -36,10 +34,60 @@ entry: define void @ti64(double %a, double %b) nounwind { entry: %tmp1 = bitcast double %a to <1 x i64> -; CHECK: movdq2q %tmp2 = bitcast double %b to <1 x i64> -; CHECK: movdq2q %tmp3 = add <1 x i64> %tmp1, %tmp2 +; CHECK: addq %rax, %rcx store <1 x i64> %tmp3, <1 x i64>* null ret void } + +; MMX intrinsics calls get us MMX instructions. + +define void @ti8a(double %a, double %b) nounwind { +entry: + %tmp1 = bitcast double %a to x86_mmx +; CHECK: movdq2q + %tmp2 = bitcast double %b to x86_mmx +; CHECK: movdq2q + %tmp3 = tail call x86_mmx @llvm.x86.mmx.padd.b(x86_mmx %tmp1, x86_mmx %tmp2) + store x86_mmx %tmp3, x86_mmx* null + ret void +} + +define void @ti16a(double %a, double %b) nounwind { +entry: + %tmp1 = bitcast double %a to x86_mmx +; CHECK: movdq2q + %tmp2 = bitcast double %b to x86_mmx +; CHECK: movdq2q + %tmp3 = tail call x86_mmx @llvm.x86.mmx.padd.w(x86_mmx %tmp1, x86_mmx %tmp2) + store x86_mmx %tmp3, x86_mmx* null + ret void +} + +define void @ti32a(double %a, double %b) nounwind { +entry: + %tmp1 = bitcast double %a to x86_mmx +; CHECK: movdq2q + %tmp2 = bitcast double %b to x86_mmx +; CHECK: movdq2q + %tmp3 = tail call x86_mmx @llvm.x86.mmx.padd.d(x86_mmx %tmp1, x86_mmx %tmp2) + store x86_mmx %tmp3, x86_mmx* null + ret void +} + +define void @ti64a(double %a, double %b) nounwind { +entry: + %tmp1 = bitcast double %a to x86_mmx +; CHECK: movdq2q + %tmp2 = bitcast double %b to x86_mmx +; CHECK: movdq2q + %tmp3 = tail call x86_mmx @llvm.x86.mmx.padd.q(x86_mmx %tmp1, x86_mmx %tmp2) + store x86_mmx %tmp3, x86_mmx* null + ret void +} + +declare x86_mmx @llvm.x86.mmx.padd.b(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padd.w(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padd.d(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padd.q(x86_mmx, x86_mmx) diff --git a/test/CodeGen/X86/fast-isel-bc.ll b/test/CodeGen/X86/fast-isel-bc.ll index 8d7dc8f9a7..db846889d8 100644 --- a/test/CodeGen/X86/fast-isel-bc.ll +++ b/test/CodeGen/X86/fast-isel-bc.ll @@ -5,15 +5,19 @@ target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128" target triple = "x86_64-apple-darwin9.8" -declare void @func2(<1 x i64>) +declare void @func2(x86_mmx) define void @func1() nounwind { ; This isn't spectacular, but it's MMX code at -O0... -; CHECK: movl $2, %eax -; CHECK: movd %rax, %mm0 -; CHECK: movd %mm0, %rdi +; CHECK: movq2dq %mm0, %xmm0 +; For now, handling of x86_mmx parameters in fast Isel is unimplemented, +; so we get pretty poor code. The below is preferable. +; CHEK: movl $2, %eax +; CHEK: movd %rax, %mm0 +; CHEK: movd %mm0, %rdi - call void @func2(<1 x i64> <i64 2>) + %tmp0 = bitcast <2 x i32><i32 0, i32 2> to x86_mmx + call void @func2(x86_mmx %tmp0) ret void } diff --git a/test/CodeGen/X86/mmx-arg-passing.ll b/test/CodeGen/X86/mmx-arg-passing.ll index 426e98e019..b348512b57 100644 --- a/test/CodeGen/X86/mmx-arg-passing.ll +++ b/test/CodeGen/X86/mmx-arg-passing.ll @@ -1,24 +1,27 @@ -; RUN: llc < %s -mtriple=i386-apple-darwin -mattr=+mmx | grep mm0 | count 3 -; RUN: llc < %s -mtriple=i386-apple-darwin -mattr=+mmx | grep esp | count 1 +; RUN: llc < %s -mtriple=i386-apple-darwin -mattr=+mmx | grep mm0 | count 1 +; RUN: llc < %s -mtriple=i386-apple-darwin -mattr=+mmx | grep esp | count 2 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | grep xmm0 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | grep rdi ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | not grep movups ; ; On Darwin x86-32, v8i8, v4i16, v2i32 values are passed in MM[0-2]. -; On Darwin x86-32, v1i64 values are passed in memory. +; On Darwin x86-32, v1i64 values are passed in memory. In this example, they +; are never moved into an MM register at all. ; On Darwin x86-64, v8i8, v4i16, v2i32 values are passed in XMM[0-7]. ; On Darwin x86-64, v1i64 values are passed in 64-bit GPRs. -@u1 = external global <8 x i8> +@u1 = external global x86_mmx -define void @t1(<8 x i8> %v1) nounwind { - store <8 x i8> %v1, <8 x i8>* @u1, align 8 +define void @t1(x86_mmx %v1) nounwind { + store x86_mmx %v1, x86_mmx* @u1, align 8 ret void } -@u2 = external global <1 x i64> +@u2 = external global x86_mmx define void @t2(<1 x i64> %v1) nounwind { - store <1 x i64> %v1, <1 x i64>* @u2, align 8 + %tmp = bitcast <1 x i64> %v1 to x86_mmx + store x86_mmx %tmp, x86_mmx* @u2, align 8 ret void } + diff --git a/test/CodeGen/X86/mmx-arg-passing2.ll b/test/CodeGen/X86/mmx-arg-passing2.ll index c42af08236..c132d311b9 100644 --- a/test/CodeGen/X86/mmx-arg-passing2.ll +++ b/test/CodeGen/X86/mmx-arg-passing2.ll @@ -1,17 +1,21 @@ -; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | grep movq2dq | count 1 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | grep movdq2q | count 2 +; Since the add is not an MMX add, we don't have a movq2dq any more. @g_v8qi = external global <8 x i8> define void @t1() nounwind { %tmp3 = load <8 x i8>* @g_v8qi, align 8 - %tmp4 = tail call i32 (...)* @pass_v8qi( <8 x i8> %tmp3 ) nounwind + %tmp3a = bitcast <8 x i8> %tmp3 to x86_mmx + %tmp4 = tail call i32 (...)* @pass_v8qi( x86_mmx %tmp3a ) nounwind ret void } -define void @t2(<8 x i8> %v1, <8 x i8> %v2) nounwind { - %tmp3 = add <8 x i8> %v1, %v2 - %tmp4 = tail call i32 (...)* @pass_v8qi( <8 x i8> %tmp3 ) nounwind +define void @t2(x86_mmx %v1, x86_mmx %v2) nounwind { + %v1a = bitcast x86_mmx %v1 to <8 x i8> + %v2b = bitcast x86_mmx %v2 to <8 x i8> + %tmp3 = add <8 x i8> %v1a, %v2b + %tmp3a = bitcast <8 x i8> %tmp3 to x86_mmx + %tmp4 = tail call i32 (...)* @pass_v8qi( x86_mmx %tmp3a ) nounwind ret void } diff --git a/test/CodeGen/X86/mmx-arith.ll b/test/CodeGen/X86/mmx-arith.ll index e4dfdbfe1b..6817487324 100644 --- a/test/CodeGen/X86/mmx-arith.ll +++ b/test/CodeGen/X86/mmx-arith.ll @@ -1,131 +1,309 @@ ; RUN: llc < %s -march=x86 -mattr=+mmx ;; A basic sanity check to make sure that MMX arithmetic actually compiles. +;; First is a straight translation of the original with bitcasts as needed. -define void @foo(<8 x i8>* %A, <8 x i8>* %B) { +define void @foo(x86_mmx* %A, x86_mmx* %B) { entry: - %tmp1 = load <8 x i8>* %A ; <<8 x i8>> [#uses=1] - %tmp3 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp4 = add <8 x i8> %tmp1, %tmp3 ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp4, <8 x i8>* %A - %tmp7 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp12 = tail call <8 x i8> @llvm.x86.mmx.padds.b( <8 x i8> %tmp4, <8 x i8> %tmp7 ) ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp12, <8 x i8>* %A - %tmp16 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp21 = tail call <8 x i8> @llvm.x86.mmx.paddus.b( <8 x i8> %tmp12, <8 x i8> %tmp16 ) ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp21, <8 x i8>* %A - %tmp27 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp28 = sub <8 x i8> %tmp21, %tmp27 ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp28, <8 x i8>* %A - %tmp31 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp36 = tail call <8 x i8> @llvm.x86.mmx.psubs.b( <8 x i8> %tmp28, <8 x i8> %tmp31 ) ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp36, <8 x i8>* %A - %tmp40 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp45 = tail call <8 x i8> @llvm.x86.mmx.psubus.b( <8 x i8> %tmp36, <8 x i8> %tmp40 ) ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp45, <8 x i8>* %A - %tmp51 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp52 = mul <8 x i8> %tmp45, %tmp51 ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp52, <8 x i8>* %A - %tmp57 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp58 = and <8 x i8> %tmp52, %tmp57 ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp58, <8 x i8>* %A - %tmp63 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp64 = or <8 x i8> %tmp58, %tmp63 ; <<8 x i8>> [#uses=2] - store <8 x i8> %tmp64, <8 x i8>* %A - %tmp69 = load <8 x i8>* %B ; <<8 x i8>> [#uses=1] - %tmp70 = xor <8 x i8> %tmp64, %tmp69 ; <<8 x i8>> [#uses=1] - store <8 x i8> %tmp70, <8 x i8>* %A + %tmp1 = load x86_mmx* %A ; <x86_mmx> [#uses=1] + %tmp3 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp1a = bitcast x86_mmx %tmp1 to <8 x i8> + %tmp3a = bitcast x86_mmx %tmp3 to <8 x i8> + %tmp4 = add <8 x i8> %tmp1a, %tmp3a ; <<8 x i8>> [#uses=2] + %tmp4a = bitcast <8 x i8> %tmp4 to x86_mmx + store x86_mmx %tmp4a, x86_mmx* %A + %tmp7 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp12 = tail call x86_mmx @llvm.x86.mmx.padds.b( x86_mmx %tmp4a, x86_mmx %tmp7 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp12, x86_mmx* %A + %tmp16 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp21 = tail call x86_mmx @llvm.x86.mmx.paddus.b( x86_mmx %tmp12, x86_mmx %tmp16 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp21, x86_mmx* %A + %tmp27 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp21a = bitcast x86_mmx %tmp21 to <8 x i8> + %tmp27a = bitcast x86_mmx %tmp27 to <8 x i8> + %tmp28 = sub <8 x i8> %tmp21a, %tmp27a ; <<8 x i8>> [#uses=2] + %tmp28a = bitcast <8 x i8> %tmp28 to x86_mmx + store x86_mmx %tmp28a, x86_mmx* %A + %tmp31 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp36 = tail call x86_mmx @llvm.x86.mmx.psubs.b( x86_mmx %tmp28a, x86_mmx %tmp31 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp36, x86_mmx* %A + %tmp40 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp45 = tail call x86_mmx @llvm.x86.mmx.psubus.b( x86_mmx %tmp36, x86_mmx %tmp40 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp45, x86_mmx* %A + %tmp51 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp45a = bitcast x86_mmx %tmp45 to <8 x i8> + %tmp51a = bitcast x86_mmx %tmp51 to <8 x i8> + %tmp52 = mul <8 x i8> %tmp45a, %tmp51a ; <<8 x i8>> [#uses=2] + %tmp52a = bitcast <8 x i8> %tmp52 to x86_mmx + store x86_mmx %tmp52a, x86_mmx* %A + %tmp57 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp57a = bitcast x86_mmx %tmp57 to <8 x i8> + %tmp58 = and <8 x i8> %tmp52, %tmp57a ; <<8 x i8>> [#uses=2] + %tmp58a = bitcast <8 x i8> %tmp58 to x86_mmx + store x86_mmx %tmp58a, x86_mmx* %A + %tmp63 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp63a = bitcast x86_mmx %tmp63 to <8 x i8> + %tmp64 = or <8 x i8> %tmp58, %tmp63a ; <<8 x i8>> [#uses=2] + %tmp64a = bitcast <8 x i8> %tmp64 to x86_mmx + store x86_mmx %tmp64a, x86_mmx* %A + %tmp69 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp69a = bitcast x86_mmx %tmp69 to <8 x i8> + %tmp64b = bitcast x86_mmx %tmp64a to <8 x i8> + %tmp70 = xor <8 x i8> %tmp64b, %tmp69a ; <<8 x i8>> [#uses=1] + %tmp70a = bitcast <8 x i8> %tmp70 to x86_mmx + store x86_mmx %tmp70a, x86_mmx* %A tail call void @llvm.x86.mmx.emms( ) ret void } -define void @baz(<2 x i32>* %A, <2 x i32>* %B) { +define void @baz(x86_mmx* %A, x86_mmx* %B) { entry: - %tmp1 = load <2 x i32>* %A ; <<2 x i32>> [#uses=1] - %tmp3 = load <2 x i32>* %B ; <<2 x i32>> [#uses=1] - %tmp4 = add <2 x i32> %tmp1, %tmp3 ; <<2 x i32>> [#uses=2] - store <2 x i32> %tmp4, <2 x i32>* %A - %tmp9 = load <2 x i32>* %B ; <<2 x i32>> [#uses=1] - %tmp10 = sub <2 x i32> %tmp4, %tmp9 ; <<2 x i32>> [#uses=2] - store <2 x i32> %tmp10, <2 x i32>* %A - %tmp15 = load <2 x i32>* %B ; <<2 x i32>> [#uses=1] - %tmp16 = mul <2 x i32> %tmp10, %tmp15 ; <<2 x i32>> [#uses=2] - store <2 x i32> %tmp16, <2 x i32>* %A - %tmp21 = load <2 x i32>* %B ; <<2 x i32>> [#uses=1] - %tmp22 = and <2 x i32> %tmp16, %tmp21 ; <<2 x i32>> [#uses=2] - store <2 x i32> %tmp22, <2 x i32>* %A - %tmp27 = load <2 x i32>* %B ; <<2 x i32>> [#uses=1] - %tmp28 = or <2 x i32> %tmp22, %tmp27 ; <<2 x i32>> [#uses=2] - store <2 x i32> %tmp28, <2 x i32>* %A - %tmp33 = load <2 x i32>* %B ; <<2 x i32>> [#uses=1] - %tmp34 = xor <2 x i32> %tmp28, %tmp33 ; <<2 x i32>> [#uses=1] - store <2 x i32> %tmp34, <2 x i32>* %A + %tmp1 = load x86_mmx* %A ; <x86_mmx> [#uses=1] + %tmp3 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp1a = bitcast x86_mmx %tmp1 to <2 x i32> + %tmp3a = bitcast x86_mmx %tmp3 to <2 x i32> + %tmp4 = add <2 x i32> %tmp1a, %tmp3a ; <<2 x i32>> [#uses=2] + %tmp4a = bitcast <2 x i32> %tmp4 to x86_mmx + store x86_mmx %tmp4a, x86_mmx* %A + %tmp9 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp9a = bitcast x86_mmx %tmp9 to <2 x i32> + %tmp10 = sub <2 x i32> %tmp4, %tmp9a ; <<2 x i32>> [#uses=2] + %tmp10a = bitcast <2 x i32> %tmp4 to x86_mmx + store x86_mmx %tmp10a, x86_mmx* %A + %tmp15 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp10b = bitcast x86_mmx %tmp10a to <2 x i32> + %tmp15a = bitcast x86_mmx %tmp15 to <2 x i32> + %tmp16 = mul <2 x i32> %tmp10b, %tmp15a ; <<2 x i32>> [#uses=2] + %tmp16a = bitcast <2 x i32> %tmp16 to x86_mmx + store x86_mmx %tmp16a, x86_mmx* %A + %tmp21 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp16b = bitcast x86_mmx %tmp16a to <2 x i32> + %tmp21a = bitcast x86_mmx %tmp21 to <2 x i32> + %tmp22 = and <2 x i32> %tmp16b, %tmp21a ; <<2 x i32>> [#uses=2] + %tmp22a = bitcast <2 x i32> %tmp22 to x86_mmx + store x86_mmx %tmp22a, x86_mmx* %A + %tmp27 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp22b = bitcast x86_mmx %tmp22a to <2 x i32> + %tmp27a = bitcast x86_mmx %tmp27 to <2 x i32> + %tmp28 = or <2 x i32> %tmp22b, %tmp27a ; <<2 x i32>> [#uses=2] + %tmp28a = bitcast <2 x i32> %tmp28 to x86_mmx + store x86_mmx %tmp28a, x86_mmx* %A + %tmp33 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp28b = bitcast x86_mmx %tmp28a to <2 x i32> + %tmp33a = bitcast x86_mmx %tmp33 to <2 x i32> + %tmp34 = xor <2 x i32> %tmp28b, %tmp33a ; <<2 x i32>> [#uses=1] + %tmp34a = bitcast <2 x i32> %tmp34 to x86_mmx + store x86_mmx %tmp34a, x86_mmx* %A tail call void @llvm.x86.mmx.emms( ) ret void } -define void @bar(<4 x i16>* %A, <4 x i16>* %B) { +define void @bar(x86_mmx* %A, x86_mmx* %B) { entry: - %tmp1 = load <4 x i16>* %A ; <<4 x i16>> [#uses=1] - %tmp3 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp4 = add <4 x i16> %tmp1, %tmp3 ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp4, <4 x i16>* %A - %tmp7 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp12 = tail call <4 x i16> @llvm.x86.mmx.padds.w( <4 x i16> %tmp4, <4 x i16> %tmp7 ) ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp12, <4 x i16>* %A - %tmp16 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp21 = tail call <4 x i16> @llvm.x86.mmx.paddus.w( <4 x i16> %tmp12, <4 x i16> %tmp16 ) ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp21, <4 x i16>* %A - %tmp27 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp28 = sub <4 x i16> %tmp21, %tmp27 ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp28, <4 x i16>* %A - %tmp31 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp36 = tail call <4 x i16> @llvm.x86.mmx.psubs.w( <4 x i16> %tmp28, <4 x i16> %tmp31 ) ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp36, <4 x i16>* %A - %tmp40 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp45 = tail call <4 x i16> @llvm.x86.mmx.psubus.w( <4 x i16> %tmp36, <4 x i16> %tmp40 ) ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp45, <4 x i16>* %A - %tmp51 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp52 = mul <4 x i16> %tmp45, %tmp51 ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp52, <4 x i16>* %A - %tmp55 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp60 = tail call <4 x i16> @llvm.x86.mmx.pmulh.w( <4 x i16> %tmp52, <4 x i16> %tmp55 ) ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp60, <4 x i16>* %A - %tmp64 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp69 = tail call <2 x i32> @llvm.x86.mmx.pmadd.wd( <4 x i16> %tmp60, <4 x i16> %tmp64 ) ; <<2 x i32>> [#uses=1] - %tmp70 = bitcast <2 x i32> %tmp69 to <4 x i16> ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp70, <4 x i16>* %A - %tmp75 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp76 = and <4 x i16> %tmp70, %tmp75 ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp76, <4 x i16>* %A - %tmp81 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp82 = or <4 x i16> %tmp76, %tmp81 ; <<4 x i16>> [#uses=2] - store <4 x i16> %tmp82, <4 x i16>* %A - %tmp87 = load <4 x i16>* %B ; <<4 x i16>> [#uses=1] - %tmp88 = xor <4 x i16> %tmp82, %tmp87 ; <<4 x i16>> [#uses=1] - store <4 x i16> %tmp88, <4 x i16>* %A + %tmp1 = load x86_mmx* %A ; <x86_mmx> [#uses=1] + %tmp3 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp1a = bitcast x86_mmx %tmp1 to <4 x i16> + %tmp3a = bitcast x86_mmx %tmp3 to <4 x i16> + %tmp4 = add <4 x i16> %tmp1a, %tmp3a ; <<4 x i16>> [#uses=2] + %tmp4a = bitcast <4 x i16> %tmp4 to x86_mmx + store x86_mmx %tmp4a, x86_mmx* %A + %tmp7 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp12 = tail call x86_mmx @llvm.x86.mmx.padds.w( x86_mmx %tmp4a, x86_mmx %tmp7 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp12, x86_mmx* %A + %tmp16 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp21 = tail call x86_mmx @llvm.x86.mmx.paddus.w( x86_mmx %tmp12, x86_mmx %tmp16 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp21, x86_mmx* %A + %tmp27 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp21a = bitcast x86_mmx %tmp21 to <4 x i16> + %tmp27a = bitcast x86_mmx %tmp27 to <4 x i16> + %tmp28 = sub <4 x i16> %tmp21a, %tmp27a ; <<4 x i16>> [#uses=2] + %tmp28a = bitcast <4 x i16> %tmp28 to x86_mmx + store x86_mmx %tmp28a, x86_mmx* %A + %tmp31 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp36 = tail call x86_mmx @llvm.x86.mmx.psubs.w( x86_mmx %tmp28a, x86_mmx %tmp31 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp36, x86_mmx* %A + %tmp40 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp45 = tail call x86_mmx @llvm.x86.mmx.psubus.w( x86_mmx %tmp36, x86_mmx %tmp40 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp45, x86_mmx* %A + %tmp51 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp45a = bitcast x86_mmx %tmp45 to <4 x i16> + %tmp51a = bitcast x86_mmx %tmp51 to <4 x i16> + %tmp52 = mul <4 x i16> %tmp45a, %tmp51a ; <<4 x i16>> [#uses=2] + %tmp52a = bitcast <4 x i16> %tmp52 to x86_mmx + store x86_mmx %tmp52a, x86_mmx* %A + %tmp55 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp60 = tail call x86_mmx @llvm.x86.mmx.pmulh.w( x86_mmx %tmp52a, x86_mmx %tmp55 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp60, x86_mmx* %A + %tmp64 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp69 = tail call x86_mmx @llvm.x86.mmx.pmadd.wd( x86_mmx %tmp60, x86_mmx %tmp64 ) ; <x86_mmx> [#uses=1] + %tmp70 = bitcast x86_mmx %tmp69 to x86_mmx ; <x86_mmx> [#uses=2] + store x86_mmx %tmp70, x86_mmx* %A + %tmp75 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp70a = bitcast x86_mmx %tmp70 to <4 x i16> + %tmp75a = bitcast x86_mmx %tmp75 to <4 x i16> + %tmp76 = and <4 x i16> %tmp70a, %tmp75a ; <<4 x i16>> [#uses=2] + %tmp76a = bitcast <4 x i16> %tmp76 to x86_mmx + store x86_mmx %tmp76a, x86_mmx* %A + %tmp81 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp76b = bitcast x86_mmx %tmp76a to <4 x i16> + %tmp81a = bitcast x86_mmx %tmp81 to <4 x i16> + %tmp82 = or <4 x i16> %tmp76b, %tmp81a ; <<4 x i16>> [#uses=2] + %tmp82a = bitcast <4 x i16> %tmp82 to x86_mmx + store x86_mmx %tmp82a, x86_mmx* %A + %tmp87 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp82b = bitcast x86_mmx %tmp82a to <4 x i16> + %tmp87a = bitcast x86_mmx %tmp87 to <4 x i16> + %tmp88 = xor <4 x i16> %tmp82b, %tmp87a ; <<4 x i16>> [#uses=1] + %tmp88a = bitcast <4 x i16> %tmp88 to x86_mmx + store x86_mmx %tmp88a, x86_mmx* %A tail call void @llvm.x86.mmx.emms( ) ret void } -declare <8 x i8> @llvm.x86.mmx.padds.b(<8 x i8>, <8 x i8>) +;; The following is modified to use MMX intrinsics everywhere they work. -declare <8 x i8> @llvm.x86.mmx.paddus.b(<8 x i8>, <8 x i8>) +define void @fooa(x86_mmx* %A, x86_mmx* %B) { +entry: + %tmp1 = load x86_mmx* %A ; <x86_mmx> [#uses=1] + %tmp3 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp4 = tail call x86_mmx @llvm.x86.mmx.padd.b( x86_mmx %tmp1, x86_mmx %tmp3 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp4, x86_mmx* %A + %tmp7 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp12 = tail call x86_mmx @llvm.x86.mmx.padds.b( x86_mmx %tmp4, x86_mmx %tmp7 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp12, x86_mmx* %A + %tmp16 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp21 = tail call x86_mmx @llvm.x86.mmx.paddus.b( x86_mmx %tmp12, x86_mmx %tmp16 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp21, x86_mmx* %A + %tmp27 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp28 = tail call x86_mmx @llvm.x86.mmx.psub.b( x86_mmx %tmp21, x86_mmx %tmp27 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp28, x86_mmx* %A + %tmp31 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp36 = tail call x86_mmx @llvm.x86.mmx.psubs.b( x86_mmx %tmp28, x86_mmx %tmp31 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp36, x86_mmx* %A + %tmp40 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp45 = tail call x86_mmx @llvm.x86.mmx.psubus.b( x86_mmx %tmp36, x86_mmx %tmp40 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp45, x86_mmx* %A + %tmp51 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp51a = bitcast x86_mmx %tmp51 to i64 + %tmp51aa = bitcast i64 %tmp51a to <8 x i8> + %tmp51b = bitcast x86_mmx %tmp45 to <8 x i8> + %tmp52 = mul <8 x i8> %tmp51b, %tmp51aa ; <x86_mmx> [#uses=2] + %tmp52a = bitcast <8 x i8> %tmp52 to i64 + %tmp52aa = bitcast i64 %tmp52a to x86_mmx + store x86_mmx %tmp52aa, x86_mmx* %A + %tmp57 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp58 = tail call x86_mmx @llvm.x86.mmx.pand( x86_mmx %tmp51, x86_mmx %tmp57 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp58, x86_mmx* %A + %tmp63 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp64 = tail call x86_mmx @llvm.x86.mmx.por( x86_mmx %tmp58, x86_mmx %tmp63 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp64, x86_mmx* %A + %tmp69 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp70 = tail call x86_mmx @llvm.x86.mmx.pxor( x86_mmx %tmp64, x86_mmx %tmp69 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp70, x86_mmx* %A + tail call void @llvm.x86.mmx.emms( ) + ret void +} -declare <8 x i8> @llvm.x86.mmx.psubs.b(<8 x i8>, <8 x i8>) +define void @baza(x86_mmx* %A, x86_mmx* %B) { +entry: + %tmp1 = load x86_mmx* %A ; <x86_mmx> [#uses=1] + %tmp3 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp4 = tail call x86_mmx @llvm.x86.mmx.padd.d( x86_mmx %tmp1, x86_mmx %tmp3 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp4, x86_mmx* %A + %tmp9 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp10 = tail call x86_mmx @llvm.x86.mmx.psub.d( x86_mmx %tmp4, x86_mmx %tmp9 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp10, x86_mmx* %A + %tmp15 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp10a = bitcast x86_mmx %tmp10 to <2 x i32> + %tmp15a = bitcast x86_mmx %tmp15 to <2 x i32> + %tmp16 = mul <2 x i32> %tmp10a, %tmp15a ; <x86_mmx> [#uses=2] + %tmp16a = bitcast <2 x i32> %tmp16 to x86_mmx + store x86_mmx %tmp16a, x86_mmx* %A + %tmp21 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp22 = tail call x86_mmx @llvm.x86.mmx.pand( x86_mmx %tmp16a, x86_mmx %tmp21 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp22, x86_mmx* %A + %tmp27 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp28 = tail call x86_mmx @llvm.x86.mmx.por( x86_mmx %tmp22, x86_mmx %tmp27 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp28, x86_mmx* %A + %tmp33 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp34 = tail call x86_mmx @llvm.x86.mmx.pxor( x86_mmx %tmp28, x86_mmx %tmp33 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp34, x86_mmx* %A + tail call void @llvm.x86.mmx.emms( ) + ret void +} -declare <8 x i8> @llvm.x86.mmx.psubus.b(<8 x i8>, <8 x i8>) +define void @bara(x86_mmx* %A, x86_mmx* %B) { +entry: + %tmp1 = load x86_mmx* %A ; <x86_mmx> [#uses=1] + %tmp3 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp4 = tail call x86_mmx @llvm.x86.mmx.padd.w( x86_mmx %tmp1, x86_mmx %tmp3 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp4, x86_mmx* %A + %tmp7 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp12 = tail call x86_mmx @llvm.x86.mmx.padds.w( x86_mmx %tmp4, x86_mmx %tmp7 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp12, x86_mmx* %A + %tmp16 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp21 = tail call x86_mmx @llvm.x86.mmx.paddus.w( x86_mmx %tmp12, x86_mmx %tmp16 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp21, x86_mmx* %A + %tmp27 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp28 = tail call x86_mmx @llvm.x86.mmx.psub.w( x86_mmx %tmp21, x86_mmx %tmp27 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp28, x86_mmx* %A + %tmp31 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp36 = tail call x86_mmx @llvm.x86.mmx.psubs.w( x86_mmx %tmp28, x86_mmx %tmp31 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp36, x86_mmx* %A + %tmp40 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp45 = tail call x86_mmx @llvm.x86.mmx.psubus.w( x86_mmx %tmp36, x86_mmx %tmp40 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp45, x86_mmx* %A + %tmp51 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp52 = tail call x86_mmx @llvm.x86.mmx.pmull.w( x86_mmx %tmp45, x86_mmx %tmp51 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp52, x86_mmx* %A + %tmp55 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp60 = tail call x86_mmx @llvm.x86.mmx.pmulh.w( x86_mmx %tmp52, x86_mmx %tmp55 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp60, x86_mmx* %A + %tmp64 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp69 = tail call x86_mmx @llvm.x86.mmx.pmadd.wd( x86_mmx %tmp60, x86_mmx %tmp64 ) ; <x86_mmx> [#uses=1] + %tmp70 = bitcast x86_mmx %tmp69 to x86_mmx ; <x86_mmx> [#uses=2] + store x86_mmx %tmp70, x86_mmx* %A + %tmp75 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp76 = tail call x86_mmx @llvm.x86.mmx.pand( x86_mmx %tmp70, x86_mmx %tmp75 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp76, x86_mmx* %A + %tmp81 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp82 = tail call x86_mmx @llvm.x86.mmx.por( x86_mmx %tmp76, x86_mmx %tmp81 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp82, x86_mmx* %A + %tmp87 = load x86_mmx* %B ; <x86_mmx> [#uses=1] + %tmp88 = tail call x86_mmx @llvm.x86.mmx.pxor( x86_mmx %tmp82, x86_mmx %tmp87 ) ; <x86_mmx> [#uses=2] + store x86_mmx %tmp88, x86_mmx* %A + tail call void @llvm.x86.mmx.emms( ) + ret void +} -declare <4 x i16> @llvm.x86.mmx.padds.w(<4 x i16>, <4 x i16>) +declare x86_mmx @llvm.x86.mmx.paddus.b(x86_mmx, x86_mmx) -declare <4 x i16> @llvm.x86.mmx.paddus.w(<4 x i16>, <4 x i16>) +declare x86_mmx @llvm.x86.mmx.psubus.b(x86_mmx, x86_mmx) -declare <4 x i16> @llvm.x86.mmx.psubs.w(<4 x i16>, <4 x i16>) +declare x86_mmx @llvm.x86.mmx.paddus.w(x86_mmx, x86_mmx) -declare <4 x i16> @llvm.x86.mmx.psubus.w(<4 x i16>, <4 x i16>) +declare x86_mmx @llvm.x86.mmx.psubus.w(x86_mmx, x86_mmx) -declare <4 x i16> @llvm.x86.mmx.pmulh.w(<4 x i16>, <4 x i16>) +declare x86_mmx @llvm.x86.mmx.pmulh.w(x86_mmx, x86_mmx) -declare <2 x i32> @llvm.x86.mmx.pmadd.wd(<4 x i16>, <4 x i16>) +declare x86_mmx @llvm.x86.mmx.pmadd.wd(x86_mmx, x86_mmx) declare void @llvm.x86.mmx.emms() + +declare x86_mmx @llvm.x86.mmx.padd.b(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padd.w(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padd.d(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padds.b(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padds.w(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padds.d(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.psubs.b(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.psubs.w(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.psubs.d(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.psub.b(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.psub.w(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.psub.d(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.pmull.w(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.pand(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.por(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.pxor(x86_mmx, x86_mmx) + diff --git a/test/CodeGen/X86/mmx-bitcast-to-i64.ll b/test/CodeGen/X86/mmx-bitcast-to-i64.ll index 1fd8f67a0c..8b1840abf6 100644 --- a/test/CodeGen/X86/mmx-bitcast-to-i64.ll +++ b/test/CodeGen/X86/mmx-bitcast-to-i64.ll @@ -1,26 +1,31 @@ ; RUN: llc < %s -march=x86-64 | grep movd | count 4 -define i64 @foo(<1 x i64>* %p) { - %t = load <1 x i64>* %p - %u = add <1 x i64> %t, %t - %s = bitcast <1 x i64> %u to i64 +define i64 @foo(x86_mmx* %p) { + %t = load x86_mmx* %p + %u = tail call x86_mmx @llvm.x86.mmx.padd.q(x86_mmx %t, x86_mmx %t) + %s = bitcast x86_mmx %u to i64 ret i64 %s } -define i64 @goo(<2 x i32>* %p) { - %t = load <2 x i32>* %p - %u = add <2 x i32> %t, %t - %s = bitcast <2 x i32> %u to i64 +define i64 @goo(x86_mmx* %p) { + %t = load x86_mmx* %p + %u = tail call x86_mmx @llvm.x86.mmx.padd.d(x86_mmx %t, x86_mmx %t) + %s = bitcast x86_mmx %u to i64 ret i64 %s } -define i64 @hoo(<4 x i16>* %p) { - %t = load <4 x i16>* %p - %u = add <4 x i16> %t, %t - %s = bitcast <4 x i16> %u to i64 +define i64 @hoo(x86_mmx* %p) { + %t = load x86_mmx* %p + %u = tail call x86_mmx @llvm.x86.mmx.padd.w(x86_mmx %t, x86_mmx %t) + %s = bitcast x86_mmx %u to i64 ret i64 %s } -define i64 @ioo(<8 x i8>* %p) { - %t = load <8 x i8>* %p - %u = add <8 x i8> %t, %t - %s = bitcast <8 x i8> %u to i64 +define i64 @ioo(x86_mmx* %p) { + %t = load x86_mmx* %p + %u = tail call x86_mmx @llvm.x86.mmx.padd.b(x86_mmx %t, x86_mmx %t) + %s = bitcast x86_mmx %u to i64 ret i64 %s } + +declare x86_mmx @llvm.x86.mmx.padd.b(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padd.w(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padd.d(x86_mmx, x86_mmx) +declare x86_mmx @llvm.x86.mmx.padd.q(x86_mmx, x86_mmx) diff --git a/test/CodeGen/X86/mmx-insert-element.ll b/test/CodeGen/X86/mmx-insert-element.ll index a063ee1d6c..9f9bd8de2d 100644 --- a/test/CodeGen/X86/mmx-insert-element.ll +++ b/test/CodeGen/X86/mmx-insert-element.ll @@ -1,7 +1,9 @@ -; RUN: llc < %s -march=x86 -mattr=+mmx | not grep movq -; RUN: llc < %s -march=x86 -mattr=+mmx | grep psllq +; RUN: llc < %s -march=x86 -mattr=+mmx,+sse | grep movq +; RUN: llc < %s -march=x86 -mattr=+mmx,+sse | grep pshufd +; This is not an MMX operation; promoted to XMM. -define <2 x i32> @qux(i32 %A) nounwind { +define x86_mmx @qux(i32 %A) nounwind { %tmp3 = insertelement <2 x i32> < i32 0, i32 undef >, i32 %A, i32 1 ; <<2 x i32>> [#uses=1] - ret <2 x i32> %tmp3 + %tmp4 = bitcast <2 x i32> %tmp3 to x86_mmx + ret x86_mmx %tmp4 } diff --git a/test/CodeGen/X86/mmx-punpckhdq.ll b/test/CodeGen/X86/mmx-punpckhdq.ll index 0af7e017b6..4818baedc4 100644 --- a/test/CodeGen/X86/mmx-punpckhdq.ll +++ b/test/CodeGen/X86/mmx-punpckhdq.ll @@ -1,4 +1,6 @@ +; RUN: llc < %s -march=x86 -mattr=+mmx | grep pextrd ; RUN: llc < %s -march=x86 -mattr=+mmx | grep punpckhdq | count 1 +; There are no MMX operations in bork; promoted to XMM. define void @bork(<1 x i64>* %x) { entry: @@ -11,4 +13,16 @@ entry: ret void } +; pork uses MMX. + +define void @pork(x86_mmx* %x) { +entry: + %tmp2 = load x86_mmx* %x ; <x86_mmx> [#uses=1] + %tmp9 = tail call x86_mmx @llvm.x86.mmx.punpckhdq (x86_mmx %tmp2, x86_mmx %tmp2) + store x86_mmx %tmp9, x86_mmx* %x + tail call void @llvm.x86.mmx.emms( ) + ret void +} + +declare x86_mmx @llvm.x86.mmx.punpckhdq(x86_mmx, x86_mmx) declare void @llvm.x86.mmx.emms() diff --git a/test/CodeGen/X86/mmx-shift.ll b/test/CodeGen/X86/mmx-shift.ll index dd0aa2ca31..bafc75444d 100644 --- a/test/CodeGen/X86/mmx-shift.ll +++ b/test/CodeGen/X86/mmx-shift.ll @@ -5,28 +5,28 @@ define i64 @t1(<1 x i64> %mm1) nounwind { entry: - %tmp6 = tail call <1 x i64> @llvm.x86.mmx.pslli.q( <1 x i64> %mm1, i32 32 ) ; <<1 x i64>> [#uses=1] - %retval1112 = bitcast <1 x i64> %tmp6 to i64 ; <i64> [#uses=1] + %tmp = bitcast <1 x i64> %mm1 to x86_mmx + %tmp6 = tail call x86_mmx @llvm.x86.mmx.pslli.q( x86_mmx %tmp, i32 32 ) ; <x86_mmx> [#uses=1] + %retval1112 = bitcast x86_mmx %tmp6 to i64 ret i64 %retval1112 } -declare <1 x i64> @llvm.x86.mmx.pslli.q(<1 x i64>, i32) nounwind readnone +declare x86_mmx @llvm.x86.mmx.pslli.q(x86_mmx, i32) nounwind readnone -define i64 @t2(<2 x i32> %mm1, <2 x i32> %mm2) nounwind { +define i64 @t2(x86_mmx %mm1, x86_mmx %mm2) nounwind { entry: - %tmp7 = tail call <2 x i32> @llvm.x86.mmx.psra.d( <2 x i32> %mm1, <2 x i32> %mm2 ) nounwind readnone ; <<2 x i32>> [#uses=1] - %retval1112 = bitcast <2 x i32> %tmp7 to i64 ; <i64> [#uses=1] + %tmp7 = tail call x86_mmx @llvm.x86.mmx.psra.d( x86_mmx %mm1, x86_mmx %mm2 ) nounwind readnone ; <x86_mmx> [#uses=1] + %retval1112 = bitcast x86_mmx %tmp7 to i64 ret i64 %retval1112 } -declare <2 x i32> @llvm.x86.mmx.psra.d(<2 x i32>, <2 x i32>) nounwind readnone +declare x86_mmx @llvm.x86.mmx.psra.d(x86_mmx, x86_mmx) nounwind readnone -define i64 @t3(<1 x i64> %mm1, i32 %bits) nounwind { +define i64 @t3(x86_mmx %mm1, i32 %bits) nounwind { entry: - %tmp6 = bitcast <1 x i64> %mm1 to <4 x i16> ; <<4 x i16>> [#uses=1] - %tmp8 = tail call <4 x i16> @llvm.x86.mmx.psrli.w( <4 x i16> %tmp6, i32 %bits ) nounwind readnone ; <<4 x i16>> [#uses=1] - %retval1314 = bitcast <4 x i16> %tmp8 to i64 ; <i64> [#uses=1] + %tmp8 = tail call x86_mmx @llvm.x86.mmx.psrli.w( x86_mmx %mm1, i32 %bits ) nounwind readnone ; <x86_mmx> [#uses=1] + %retval1314 = bitcast x86_mmx %tmp8 to i64 ret i64 %retval1314 } -declare <4 x i16> @llvm.x86.mmx.psrli.w(<4 x i16>, i32) nounwind readnone +declare x86_mmx @llvm.x86.mmx.psrli.w(x86_mmx, i32) nounwind readnone diff --git a/test/CodeGen/X86/mmx-shuffle.ll b/test/CodeGen/X86/mmx-shuffle.ll index e3125c7345..9f7501eb7c 100644 --- a/test/CodeGen/X86/mmx-shuffle.ll +++ b/test/CodeGen/X86/mmx-shuffle.ll @@ -22,8 +22,10 @@ entry: %tmp542 = bitcast <2 x i32> %tmp529 to <4 x i16> ; <<4 x i16>> [#uses=1] %tmp543 = add <4 x i16> %tmp542, < i16 0, i16 16448, i16 24672, i16 28784 > ; <<4 x i16>> [#uses=1] %tmp555 = bitcast <4 x i16> %tmp543 to <8 x i8> ; <<8 x i8>> [#uses=1] - tail call void @llvm.x86.mmx.maskmovq( <8 x i8> zeroinitializer, <8 x i8> %tmp555, i8* null ) + %tmp556 = bitcast <8 x i8> %tmp555 to x86_mmx + %tmp557 = bitcast <8 x i8> zeroinitializer to x86_mmx + tail call void @llvm.x86.mmx.maskmovq( x86_mmx %tmp557, x86_mmx %tmp556, i8* null ) ret void } -declare void @llvm.x86.mmx.maskmovq(<8 x i8>, <8 x i8>, i8*) +declare void @llvm.x86.mmx.maskmovq(x86_mmx, x86_mmx, i8*) diff --git a/test/CodeGen/X86/mmx-vzmovl-2.ll b/test/CodeGen/X86/mmx-vzmovl-2.ll index 8253c20032..4ad420b37c 100644 --- a/test/CodeGen/X86/mmx-vzmovl-2.ll +++ b/test/CodeGen/X86/mmx-vzmovl-2.ll @@ -4,7 +4,7 @@ %struct.vS1024 = type { [8 x <4 x i32>] } %struct.vS512 = type { [4 x <4 x i32>] } -declare <1 x i64> @llvm.x86.mmx.psrli.q(<1 x i64>, i32) nounwind readnone +declare x86_mmx @llvm.x86.mmx.psrli.q(x86_mmx, i32) nounwind readnone define void @t() nounwind { entry: @@ -12,14 +12,18 @@ entry: bb554: ; preds = %bb554, %entry %sum.0.reg2mem.0 = phi <1 x i64> [ %tmp562, %bb554 ], [ zeroinitializer, %entry ] ; <<1 x i64>> [#uses=1] - %0 = load <1 x i64>* null, align 8 ; <<1 x i64>> [#uses=2] - %1 = bitcast <1 x i64> %0 to <2 x i32> ; <<2 x i32>> [#uses=1] + %0 = load x86_mmx* null, align 8 ; <<1 x i64>> [#uses=2] + %1 = bitcast x86_mmx %0 to <2 x i32> ; <<2 x i32>> [#uses=1] %tmp555 = and <2 x i32> %1, < i32 -1, i32 0 > ; <<2 x i32>> [#uses=1] - %2 = bitcast <2 x i32> %tmp555 to <1 x i64> ; <<1 x i64>> [#uses=1] - %3 = call <1 x i64> @llvm.x86.mmx.psrli.q(<1 x i64> %0, i32 32) nounwind readnone ; <<1 x i64>> [#uses=1] + %2 = bitcast <2 x i32> %tmp555 to x86_mmx ; <<1 x i64>> [#uses=1] + %3 = call x86_mmx @llvm.x86.mmx.psrli.q(x86_mmx %0, i32 32) nounwind readnone ; <<1 x i64>> [#uses=1] store <1 x i64> %sum.0.reg2mem.0, <1 x i64>* null - %tmp558 = add <1 x i64> %sum.0.reg2mem.0, %2 ; <<1 x i64>> [#uses=1] - %4 = call <1 x i64> @llvm.x86.mmx.psrli.q(<1 x i64> %tmp558, i32 32) nounwind readnone ; <<1 x i64>> [#uses=1] - %tmp562 = add <1 x i64> %4, %3 ; <<1 x i64>> [#uses=1] + %tmp3 = bitcast x86_mmx %2 to <1 x i64> + %tmp558 = add <1 x i64> %sum.0.reg2mem.0, %tmp3 ; <<1 x i64>> [#uses=1] + %tmp5 = bitcast <1 x i64> %tmp558 to x86_mmx + %4 = call x86_mmx @llvm.x86.mmx.psrli.q(x86_mmx %tmp5, i32 32) nounwind readnone ; <<1 x i64>> [#uses=1] + %tmp6 = bitcast x86_mmx %4 to <1 x i64> + %tmp7 = bitcast x86_mmx %3 to <1 x i64> + %tmp562 = add <1 x i64> %tmp6, %tmp7 ; <<1 x i64>> [#uses=1] br label %bb554 } diff --git a/test/CodeGen/X86/mmx-vzmovl.ll b/test/CodeGen/X86/mmx-vzmovl.ll index d21e240488..e8b34263c6 100644 --- a/test/CodeGen/X86/mmx-vzmovl.ll +++ b/test/CodeGen/X86/mmx-vzmovl.ll @@ -1,5 +1,5 @@ -; RUN: llc < %s -march=x86-64 -mattr=+mmx | grep movd -; RUN: llc < %s -march=x86-64 -mattr=+mmx | grep movq +; RUN: llc < %s -march=x86-64 -mattr=+mmx | grep movq | count 2 +; There are no MMX operations here; this is promoted to XMM. define void @foo(<1 x i64>* %a, <1 x i64>* %b) nounwind { entry: diff --git a/test/CodeGen/X86/vec_insert-5.ll b/test/CodeGen/X86/vec_insert-5.ll index 291fc0454c..471cc1611f 100644 --- a/test/CodeGen/X86/vec_insert-5.ll +++ b/test/CodeGen/X86/vec_insert-5.ll @@ -1,15 +1,16 @@ ; RUN: llc < %s -march=x86 -mattr=+sse2 > %t -; RUN: grep psllq %t | grep 32 +; RUN: grep shll %t | grep 12 ; RUN: grep pslldq %t | grep 12 ; RUN: grep psrldq %t | grep 8 ; RUN: grep psrldq %t | grep 12 +; There are no MMX operations in @t1 -define void @t1(i32 %a, <1 x i64>* %P) nounwind { +define void @t1(i32 %a, x86_mmx* %P) nounwind { %tmp12 = shl i32 %a, 12 %tmp21 = insertelement <2 x i32> undef, i32 %tmp12, i32 1 %tmp22 = insertelement <2 x i32> %tmp21, i32 0, i32 0 - %tmp23 = bitcast <2 x i32> %tmp22 to <1 x i64> - store <1 x i64> %tmp23, <1 x i64>* %P + %tmp23 = bitcast <2 x i32> %tmp22 to x86_mmx + store x86_mmx %tmp23, x86_mmx* %P ret void } diff --git a/test/CodeGen/X86/vec_insert-7.ll b/test/CodeGen/X86/vec_insert-7.ll index 9ede10f63d..ea7f919304 100644 --- a/test/CodeGen/X86/vec_insert-7.ll +++ b/test/CodeGen/X86/vec_insert-7.ll @@ -1,8 +1,12 @@ -; RUN: llc < %s -march=x86 -mattr=+mmx -mtriple=i686-apple-darwin9 -o - | grep punpckldq +; RUN: llc < %s -march=x86 -mattr=+mmx,+sse -mtriple=i686-apple-darwin9 -o - | grep pinsrd | count 2 +; MMX insertelement is not available; these are promoted to XMM. +; (Without SSE they are split to two ints, and the code is much better.) -define <2 x i32> @mmx_movzl(<2 x i32> %x) nounwind { +define x86_mmx @mmx_movzl(x86_mmx %x) nounwind { entry: - %tmp3 = insertelement <2 x i32> %x, i32 32, i32 0 ; <<2 x i32>> [#uses=1] + %tmp = bitcast x86_mmx %x to <2 x i32> + %tmp3 = insertelement <2 x i32> %tmp, i32 32, i32 0 ; <<2 x i32>> [#uses=1] %tmp8 = insertelement <2 x i32> %tmp3, i32 0, i32 1 ; <<2 x i32>> [#uses=1] - ret <2 x i32> %tmp8 + %tmp9 = bitcast <2 x i32> %tmp8 to x86_mmx + ret x86_mmx %tmp9 } diff --git a/test/CodeGen/X86/vec_zero_cse.ll b/test/CodeGen/X86/vec_zero_cse.ll index 3b15d4cc40..8aa50945e6 100644 --- a/test/CodeGen/X86/vec_zero_cse.ll +++ b/test/CodeGen/X86/vec_zero_cse.ll @@ -1,5 +1,6 @@ -; RUN: llc < %s -relocation-model=static -march=x86 -mcpu=yonah | grep pxor | count 2 -; RUN: llc < %s -relocation-model=static -march=x86 -mcpu=yonah | grep pcmpeqd | count 2 +; RUN: llc < %s -relocation-model=static -march=x86 -mcpu=yonah | grep pxor | count 1 +; RUN: llc < %s -relocation-model=static -march=x86 -mcpu=yonah | grep pcmpeqd | count 1 +; 64-bit stores here do not use MMX. @M1 = external global <1 x i64> @M2 = external global <2 x i32> |