[llvm] [NFC][AMDGPU] Pre-commit tests for IR variant - isFMAFasterThanFMulAdd (PR #121925)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Tue Jan 7 04:47:44 PST 2025
================
@@ -0,0 +1,447 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes='simplifycfg<hoist-common-insts>' -mtriple=amdgcn-- --fp-contract=fast -mcpu=gfx1030 < %s | FileCheck -check-prefix=GFX -check-prefix=FP-CONTRACT-FAST %s
+; RUN: opt -S -passes='simplifycfg<hoist-common-insts>' -mtriple=amdgcn-- --fp-contract=off --enable-unsafe-fp-math -mcpu=gfx1030 < %s | FileCheck -check-prefix=GFX -check-prefix=UNSAFE-FP-MATH %s
+; RUN: opt -S -passes='simplifycfg<hoist-common-insts>' -mtriple=amdgcn-- --fp-contract=off -mcpu=gfx1030 < %s | FileCheck -check-prefix=GFX -check-prefix=NO-UNSAFE-FP-MATH %s
+
+define double @is_profitable_f64_contract(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #0 {
+; GFX-LABEL: define double @is_profitable_f64_contract(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0:[0-9]+]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load double, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
+; GFX-NEXT: [[X:%.*]] = load double, ptr [[PTR_X]], align 8
+; GFX-NEXT: [[A_1:%.*]] = load double, ptr [[PTR_A]], align 8
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi double [ [[ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret double [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[MUL:%.*]] = fmul contract double [[X]], [[A_1]]
+; GFX-NEXT: [[ADD]] = fadd contract double 1.000000e+00, [[MUL]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[MUL1:%.*]] = fmul contract double [[X]], [[A_1]]
+; GFX-NEXT: [[SUB]] = fsub contract double [[MUL1]], [[Y]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load double, ptr %ptr_y, align 8
+ %cmp = fcmp oeq double %y, 0.000000e+00
+ %x = load double, ptr %ptr_x, align 8
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then: ; preds = %entry
+ %a_1 = load double, ptr %ptr_a, align 8
+ %mul = fmul contract double %x, %a_1
+ %add = fadd contract double 1.000000e+00, %mul
+ ret double %add
+
+if.else: ; preds = %entry
+ %a_2 = load double, ptr %ptr_a, align 8
+ %mul1 = fmul contract double %x, %a_2
+ %sub = fsub contract double %mul1, %y
+ ret double %sub
+}
+
+define double @is_profitable_f64_modifiers(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #0 {
+; GFX-LABEL: define double @is_profitable_f64_modifiers(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load double, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
+; GFX-NEXT: [[X:%.*]] = load double, ptr [[PTR_X]], align 8
+; GFX-NEXT: [[A_1:%.*]] = load double, ptr [[PTR_A]], align 8
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi double [ [[FNEG:%.*]], %[[IF_THEN]] ], [ [[FABS:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret double [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[MUL:%.*]] = fmul contract double [[X]], [[A_1]]
+; GFX-NEXT: [[ADD:%.*]] = fadd contract double 1.000000e+00, [[MUL]]
+; GFX-NEXT: [[FNEG]] = fneg double [[ADD]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[MUL1:%.*]] = fmul contract double [[X]], [[A_1]]
+; GFX-NEXT: [[SUB:%.*]] = fsub contract double [[MUL1]], [[Y]]
+; GFX-NEXT: [[FABS]] = call double @llvm.fabs.f64(double [[SUB]])
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load double, ptr %ptr_y, align 8
+ %cmp = fcmp oeq double %y, 0.000000e+00
+ %x = load double, ptr %ptr_x, align 8
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then: ; preds = %entry
+ %a_1 = load double, ptr %ptr_a, align 8
+ %mul = fmul contract double %x, %a_1
+ %add = fadd contract double 1.000000e+00, %mul
+ %fneg = fneg double %add
+ ret double %fneg
+
+if.else: ; preds = %entry
+ %a_2 = load double, ptr %ptr_a, align 8
+ %mul1 = fmul contract double %x, %a_2
+ %sub = fsub contract double %mul1, %y
+ %fabs = call double @llvm.fabs.f64(double %sub)
+ ret double %fabs
+}
+
+define float @is_profitable_f32(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #0 {
+; GFX-LABEL: define float @is_profitable_f32(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load float, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq float [[Y]], 0.000000e+00
+; GFX-NEXT: [[X:%.*]] = load float, ptr [[PTR_X]], align 8
+; GFX-NEXT: [[A_1:%.*]] = load float, ptr [[PTR_A]], align 8
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi float [ [[MUL:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret float [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[MUL]] = fmul contract float [[X]], [[A_1]]
+; GFX-NEXT: [[ADD:%.*]] = fadd contract float 1.000000e+00, [[MUL]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[MUL1:%.*]] = fmul contract float [[X]], [[A_1]]
+; GFX-NEXT: [[SUB]] = fsub contract float [[MUL1]], [[Y]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load float, ptr %ptr_y, align 8
+ %cmp = fcmp oeq float %y, 0.000000e+00
+ %x = load float, ptr %ptr_x, align 8
+ br i1 %cmp, label %if.then, label %if.else
+
+
+if.then: ; preds = %entry
+ %a_1 = load float, ptr %ptr_a, align 8
+ %mul = fmul contract float %x, %a_1
+ %add = fadd contract float 1.000000e+00, %mul
+ ret float %mul
+
+if.else: ; preds = %entry
+ %a_2 = load float, ptr %ptr_a, align 8
+ %mul1 = fmul contract float %x, %a_2
+ %sub = fsub contract float %mul1, %y
+ ret float %sub
+}
+
+define half @is_profitable_f16_preserve(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #0 {
+; GFX-LABEL: define half @is_profitable_f16_preserve(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load half, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq half [[Y]], 0xH0000
+; GFX-NEXT: [[X:%.*]] = load half, ptr [[PTR_X]], align 8
+; GFX-NEXT: [[A_1:%.*]] = load half, ptr [[PTR_A]], align 8
+; GFX-NEXT: [[MUL:%.*]] = fmul contract half [[X]], [[A_1]]
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi half [ [[MUL]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret half [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[ADD:%.*]] = fadd contract half [[Y]], [[MUL]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[SUB]] = fsub contract half [[MUL]], [[Y]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load half, ptr %ptr_y, align 8
+ %cmp = fcmp oeq half %y, 0.000000e+00
+ %x = load half, ptr %ptr_x, align 8
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then: ; preds = %entry
+ %a_1 = load half, ptr %ptr_a, align 8
+ %mul = fmul contract half %x, %a_1
+ %add = fadd contract half %y, %mul
+ ret half %mul
+
+if.else: ; preds = %entry
+ %a_2 = load half, ptr %ptr_a, align 8
+ %mul1 = fmul contract half %x, %a_2
+ %sub = fsub contract half %mul1, %y
+ ret half %sub
+}
+
+define half @is_profitable_f16_ieee(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #1 {
+; GFX-LABEL: define half @is_profitable_f16_ieee(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR1:[0-9]+]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load half, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq half [[Y]], 0xH0000
+; GFX-NEXT: [[X:%.*]] = load half, ptr [[PTR_X]], align 8
+; GFX-NEXT: [[A_1:%.*]] = load half, ptr [[PTR_A]], align 8
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi half [ [[MUL:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret half [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[MUL]] = fmul contract half [[X]], [[A_1]]
+; GFX-NEXT: [[ADD:%.*]] = fadd contract half [[Y]], [[MUL]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[MUL1:%.*]] = fmul contract half [[X]], [[A_1]]
+; GFX-NEXT: [[SUB]] = fsub contract half [[MUL1]], [[Y]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load half, ptr %ptr_y, align 8
+ %cmp = fcmp oeq half %y, 0.000000e+00
+ %x = load half, ptr %ptr_x, align 8
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then: ; preds = %entry
+ %a_1 = load half, ptr %ptr_a, align 8
+ %mul = fmul contract half %x, %a_1
+ %add = fadd contract half %y, %mul
+ ret half %mul
+
+if.else: ; preds = %entry
+ %a_2 = load half, ptr %ptr_a, align 8
+ %mul1 = fmul contract half %x, %a_2
+ %sub = fsub contract half %mul1, %y
+ ret half %sub
+}
+
+define bfloat @is_profitable_bfloat_preserve(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #0 {
+; GFX-LABEL: define bfloat @is_profitable_bfloat_preserve(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load bfloat, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq bfloat [[Y]], 0xR0000
+; GFX-NEXT: [[X:%.*]] = load bfloat, ptr [[PTR_X]], align 8
+; GFX-NEXT: [[A_1:%.*]] = load bfloat, ptr [[PTR_A]], align 8
+; GFX-NEXT: [[MUL:%.*]] = fmul contract bfloat [[X]], [[A_1]]
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi bfloat [ [[MUL]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret bfloat [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[ADD:%.*]] = fadd contract bfloat 0xR3F80, [[MUL]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[SUB]] = fsub contract bfloat [[MUL]], [[Y]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load bfloat, ptr %ptr_y, align 8
+ %cmp = fcmp oeq bfloat %y, 0.000000e+00
+ %x = load bfloat, ptr %ptr_x, align 8
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then: ; preds = %entry
+ %a_1 = load bfloat, ptr %ptr_a, align 8
+ %mul = fmul contract bfloat %x, %a_1
+ %add = fadd contract bfloat 1.000000e+00, %mul
+ ret bfloat %mul
+
+if.else: ; preds = %entry
+ %a_2 = load bfloat, ptr %ptr_a, align 8
+ %mul1 = fmul contract bfloat %x, %a_2
+ %sub = fsub contract bfloat %mul1, %y
+ ret bfloat %sub
+}
+
+define bfloat @is_profitable_bfloat_ieee(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #1 {
+; GFX-LABEL: define bfloat @is_profitable_bfloat_ieee(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR1]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load bfloat, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq bfloat [[Y]], 0xR0000
+; GFX-NEXT: [[X:%.*]] = load bfloat, ptr [[PTR_X]], align 8
+; GFX-NEXT: [[A_1:%.*]] = load bfloat, ptr [[PTR_A]], align 8
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi bfloat [ [[MUL:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret bfloat [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[MUL]] = fmul contract bfloat [[X]], [[A_1]]
+; GFX-NEXT: [[ADD:%.*]] = fadd contract bfloat 0xR3F80, [[MUL]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[MUL1:%.*]] = fmul contract bfloat [[X]], [[A_1]]
+; GFX-NEXT: [[SUB]] = fsub contract bfloat [[MUL1]], [[Y]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load bfloat, ptr %ptr_y, align 8
+ %cmp = fcmp oeq bfloat %y, 0.000000e+00
+ %x = load bfloat, ptr %ptr_x, align 8
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then: ; preds = %entry
+ %a_1 = load bfloat, ptr %ptr_a, align 8
+ %mul = fmul contract bfloat %x, %a_1
+ %add = fadd contract bfloat 1.000000e+00, %mul
+ ret bfloat %mul
+
+if.else: ; preds = %entry
+ %a_2 = load bfloat, ptr %ptr_a, align 8
+ %mul1 = fmul contract bfloat %x, %a_2
+ %sub = fsub contract bfloat %mul1, %y
+ ret bfloat %sub
+}
+
+define double @is_profitable_constant(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #1 {
+; GFX-LABEL: define double @is_profitable_constant(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR1]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load double, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
+; GFX-NEXT: [[X:%.*]] = load double, ptr [[PTR_X]], align 8
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi double [ [[ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret double [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[MUL:%.*]] = fmul contract double 2.000000e+00, [[X]]
+; GFX-NEXT: [[ADD]] = fadd contract double 1.000000e+00, [[MUL]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[MUL1:%.*]] = fmul contract double 3.000000e+00, [[X]]
+; GFX-NEXT: [[SUB]] = fsub contract double [[MUL1]], [[Y]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load double, ptr %ptr_y, align 8
+ %cmp = fcmp oeq double %y, 0.000000e+00
+ %x = load double, ptr %ptr_x, align 8
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then: ; preds = %entry
+ %mul = fmul contract double 2.000000e+00, %x
+ %add = fadd contract double 1.000000e+00, %mul
+ ret double %add
+
+if.else: ; preds = %entry
+ %mul1 = fmul contract double 3.000000e+00, %x
+ %sub = fsub contract double %mul1, %y
+ ret double %sub
+}
+
+%vec_type = type <8 x double>
+ at v1_ptr = external addrspace(3) global ptr
+ at v2_ptr = external addrspace(3) global ptr
+
+define %vec_type @is_profitable_vector(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) {
+; GFX-LABEL: define <8 x double> @is_profitable_vector(
+; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR2:[0-9]+]] {
+; GFX-NEXT: [[ENTRY:.*:]]
+; GFX-NEXT: [[Y:%.*]] = load double, ptr [[PTR_Y]], align 8
+; GFX-NEXT: [[X:%.*]] = load <8 x double>, ptr [[PTR_X]], align 8
+; GFX-NEXT: [[V1:%.*]] = load <8 x double>, ptr addrspace(3) @v1_ptr, align 64
+; GFX-NEXT: [[V2:%.*]] = load <8 x double>, ptr addrspace(3) @v2_ptr, align 64
+; GFX-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
+; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
+; GFX: [[COMMON_RET:.*]]:
+; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi <8 x double> [ [[ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
+; GFX-NEXT: ret <8 x double> [[COMMON_RET_OP]]
+; GFX: [[IF_THEN]]:
+; GFX-NEXT: [[MUL:%.*]] = fmul contract <8 x double> [[V1]], [[X]]
+; GFX-NEXT: [[ADD]] = fadd contract <8 x double> [[V2]], [[MUL]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+; GFX: [[IF_ELSE]]:
+; GFX-NEXT: [[MUL1:%.*]] = fmul contract <8 x double> [[V1]], [[X]]
+; GFX-NEXT: [[SUB]] = fsub contract <8 x double> [[MUL1]], [[V2]]
+; GFX-NEXT: br label %[[COMMON_RET]]
+;
+entry:
+ %y = load double, ptr %ptr_y, align 8
+ %x = load %vec_type, ptr %ptr_x, align 8
+ %v1 = load %vec_type, ptr addrspace(3) @v1_ptr
+ %v2 = load %vec_type, ptr addrspace(3) @v2_ptr
+ %cmp = fcmp oeq double %y, 0.000000e+00
+ br i1 %cmp, label %if.then, label %if.else
+
+if.then: ; preds = %entry
+ %mul = fmul contract %vec_type %v1, %x
----------------
arsenm wrote:
Use `<8 x double>` directly?
Also the half vectors are the most interesting cases
https://github.com/llvm/llvm-project/pull/121925
More information about the llvm-commits
mailing list